Skip to navigation

Threat reportThreat IntelligenceTL-2026-6430

AI-Assisted Malware: Halcyon Analysis of Real-World vs Lab-Only Samples (Unit 42: 12 of 405 AI-linked samples seen in production)

lowMONITORING

Halcyon's 2026-10-06 analysis, drawing on Unit 42's study of 405 AI-linked malware samples, notes that only 12 (~3%) appeared on protected production endpoints and all were caught by existing controls. AI speeds malware development and lowers skill barriers (Anthropic, Google GTIG cases) but has not produced a new intrusion method; every attack stage remains detectable. First disclosed .

Key facts

IdentifiersTL-2026-6430
Also known asAI malware measurement problem, AI-enabled malware: hype versus reality
SeverityLOW
First disclosed
Affected productsVarious Windows endpoints, cloud environments and developer/CI-CD credentials targeted by AI-assisted tooling
MalwareBroomstick, DUSTMAKER, FunkSec, Recipe Lister, Rhadamanthys, SANDCLOCK
ToolsClaude Code, Markdown instruction-file agent playbooks
MITRE ATT&CK techniquesT1027 Obfuscated Files or Information, T1036 Masquerading, T1059.007 JavaScript, T1195.002 Compromise Software Supply Chain, T1204.002 Malicious File, T1490 Inhibit System Recovery, T1491.001 Internal Defacement, T1546.015 Component Object Model Hijacking, T1553.002 Code Signing, T1555 Credentials from Password Stores, T1574.001 DLL, T1595.002 Vulnerability Scanning and 2 more in the attack flow
MotivationFINANCIAL
Target sectorshealth, government administration, education, technology, emergency services, finance
Target regionsGlobal
Detection rules9 (1 shown free)
Sources10 cited sources

Detection coverage

As of 2026-11-19, Threadlinqs Intelligence publishes 9 detection rule(s) for TL-2026-6430 across Splunk SPL, Microsoft KQL and Sigma, covering 26 indicator(s) of compromise. The whole corpus is readable without an account. Every threat comes with one complete detection in Splunk SPL, Microsoft KQL and Sigma, free for everyone; a free account adds daily TL Creds to unlock the rest. Paid plans add a larger daily allowance, correlation and the MCP server. Threadlinqs MCP server · View plans.

Free detection: AI-Linked Malware Known-SHA256 Execution or Revoked Signer (FunkSec, Recipe Lister, Oyster, Rhadamanthys, 360Util.dll) — Initial Access

Fires when a process start or image load matches one of the 12 Unit 42 SHA256 indicators for the AI-linked samples seen in production (FunkSec ransomware, Recipe Lister trojanized NSIS installer, Oyster/CleanBoost fake Dropbox installer, Rhadamanthys redist.exe, 360Util.dll COM-hijack DLL), or when a binary signed by the since-revoked 'Global Tech Allies Ltd' certificate runs. Hash and signer matches are high-fidelity; AI authorship changes how the code is written, not how it executes, so classic IOC and signer-rarity detection still applies.

  • Languages: Splunk SPL, Microsoft KQL, Sigma
  • ATT&CK: T1204.002, T1195.002, T1553.002, T1036
  • Data sources: sysmon, microsoft_defender_for_endpoint
  • Severity: high
  • Confidence: high

Splunk SPL

index=sysmon sourcetype="XmlWinEventLog:Microsoft-Windows-Sysmon/Operational" (EventCode=1 OR EventCode=7) (Hashes IN ("*1619bcad3785be31ac2fdee0ab91392d08d9392032246e42673c3cb8964d4cb7*","*5226ea8e0f516565ba825a1bbed10020982c16414750237068b602c5b4ac6abd*","*dcf536edd67a98868759f4e72bcbd1f4404c70048a2a3257e77d8af06cb036ac*","*66dbf939c00b09d8d22c692864b68c4a602e7a59c4b925b2e2bef57b1ad047bd*","*c233aec7917cf34294c19dd60ff79a6e0fac5ed6f0cb57af98013c08201a7a1c*","*e622f3b743c7fc0a011b07a2e656aa2b5e50a4876721bcf1f405d582ca4cda22*","*b1ef7b267d887e34bf0242a94b38e7dc9fd5e6f8b2c5c440ce4ec98cc74642fb*","*20ed21bfdb7aa970b12e7368eba8e26a711752f1cc5416b6fd6629d0e2a44e5d*","*dd15ce869aa79884753e3baad19b0437075202be86268b84f3ec2303e1ecd966*","*c398b3e06ef860670b9597daed85632834fa961aea87164b8ba8bb2f094a14ef*","*bb932056cae8940742e50b4f2b994a802e703f7bc235e7dd647d085ae2b2baf7*","*4fb58687a364c3f6d6f7e0ca03654f9dec0f8832a499d61d40b0d424db1b1b14*") OR Signature="Global Tech Allies Ltd")
| eval ioc_match=if(like(Signature,"Global Tech Allies%"),"revoked_signer_Global_Tech_Allies_Ltd","sha256_ioc_unit42_ai_linked")
| stats count min(_time) as first_seen max(_time) as last_seen values(ioc_match) as ioc_match values(Image) as image values(ImageLoaded) as image_loaded values(ParentImage) as parent_image values(CommandLine) as command_line values(Hashes) as hashes by host
| convert ctime(first_seen) ctime(last_seen)

Microsoft KQL

let lookback = 24h;
let ioc_sha256 = dynamic(["1619bcad3785be31ac2fdee0ab91392d08d9392032246e42673c3cb8964d4cb7", "5226ea8e0f516565ba825a1bbed10020982c16414750237068b602c5b4ac6abd", "dcf536edd67a98868759f4e72bcbd1f4404c70048a2a3257e77d8af06cb036ac", "66dbf939c00b09d8d22c692864b68c4a602e7a59c4b925b2e2bef57b1ad047bd", "c233aec7917cf34294c19dd60ff79a6e0fac5ed6f0cb57af98013c08201a7a1c", "e622f3b743c7fc0a011b07a2e656aa2b5e50a4876721bcf1f405d582ca4cda22", "b1ef7b267d887e34bf0242a94b38e7dc9fd5e6f8b2c5c440ce4ec98cc74642fb", "20ed21bfdb7aa970b12e7368eba8e26a711752f1cc5416b6fd6629d0e2a44e5d", "dd15ce869aa79884753e3baad19b0437075202be86268b84f3ec2303e1ecd966", "c398b3e06ef860670b9597daed85632834fa961aea87164b8ba8bb2f094a14ef", "bb932056cae8940742e50b4f2b994a802e703f7bc235e7dd647d085ae2b2baf7", "4fb58687a364c3f6d6f7e0ca03654f9dec0f8832a499d61d40b0d424db1b1b14"]);
let hash_hits = union withsource=SourceTable DeviceProcessEvents, DeviceFileEvents, DeviceImageLoadEvents
    | where Timestamp > ago(lookback)
    | where SHA256 in~ (ioc_sha256)
    | project Timestamp, DeviceName, Detection = strcat("sha256_ioc_", SourceTable), FileName, FolderPath, FileHash = SHA256, ParentProcess = InitiatingProcessFileName, ParentCommandLine = InitiatingProcessCommandLine;
let signer_hits = DeviceFileCertificateInfo
    | where Timestamp > ago(lookback)
    | where Signer =~ "Global Tech Allies Ltd"
    | project Timestamp, DeviceName, Detection = "revoked_signer_global_tech_allies", FileName = "", FolderPath = "", FileHash = SHA1, ParentProcess = "", ParentCommandLine = "";
union hash_hits, signer_hits
| order by Timestamp desc

Sigma

title: TL-2026-6430 Known AI-Linked Malware SHA256 Execution (FunkSec, Recipe Lister, Oyster, Rhadamanthys, 360Util)
id: 6f1b2c4e-8a3d-4e57-9c10-5d2a7b9e4f61
status: experimental
description: Detects execution of binaries whose SHA256 matches the Unit 42 indicators for AI-linked malware observed on protected production endpoints (FunkSec ransomware, Recipe Lister NSIS installer, Oyster/CleanBoost fake Dropbox installer, Rhadamanthys redist.exe, 360Util.dll host). Pair with signer-rarity monitoring for the revoked 'Global Tech Allies Ltd' certificate.
references:
  - https://unit42.paloaltonetworks.com/ai-enabled-malware-analysis/
  - https://www.halcyon.ai/blog/ai-malware-panic-measurement-problem
author: Threadlinqs Team
date: 2026-11-19
tags:
  - attack.initial-access
  - attack.t1195.002
  - attack.execution
  - attack.t1204.002
  - attack.defense-evasion
  - attack.t1553.002
  - attack.t1036
logsource:
  category: process_creation
  product: windows
detection:
  selection:
    Hashes|contains:
      - 'SHA256=1619bcad3785be31ac2fdee0ab91392d08d9392032246e42673c3cb8964d4cb7'
      - 'SHA256=5226ea8e0f516565ba825a1bbed10020982c16414750237068b602c5b4ac6abd'
      - 'SHA256=dcf536edd67a98868759f4e72bcbd1f4404c70048a2a3257e77d8af06cb036ac'
      - 'SHA256=66dbf939c00b09d8d22c692864b68c4a602e7a59c4b925b2e2bef57b1ad047bd'
      - 'SHA256=c233aec7917cf34294c19dd60ff79a6e0fac5ed6f0cb57af98013c08201a7a1c'
      - 'SHA256=e622f3b743c7fc0a011b07a2e656aa2b5e50a4876721bcf1f405d582ca4cda22'
      - 'SHA256=b1ef7b267d887e34bf0242a94b38e7dc9fd5e6f8b2c5c440ce4ec98cc74642fb'
      - 'SHA256=20ed21bfdb7aa970b12e7368eba8e26a711752f1cc5416b6fd6629d0e2a44e5d'
      - 'SHA256=dd15ce869aa79884753e3baad19b0437075202be86268b84f3ec2303e1ecd966'
      - 'SHA256=c398b3e06ef860670b9597daed85632834fa961aea87164b8ba8bb2f094a14ef'
      - 'SHA256=bb932056cae8940742e50b4f2b994a802e703f7bc235e7dd647d085ae2b2baf7'
      - 'SHA256=4fb58687a364c3f6d6f7e0ca03654f9dec0f8832a499d61d40b0d424db1b1b14'
  condition: selection
falsepositives:
  - Malware analysts or sandboxes intentionally detonating the published samples
level: high

Known false positives:

  • Malware analysts or sandbox/detonation hosts deliberately running the published Unit 42 samples
  • Threat-intel test systems that replay known malware hashes for detection validation

Tuning notes: Scope to endpoints by excluding documented sandbox/analyst hosts (host lookup). Refresh the SHA256 list when Unit 42 or Halcyon publish new samples. Ensure Sysmon logs SHA256 in the Hashes field (HashType=SHA256) and that Event ID 7 image-load logging is enabled for the signer check. Adjust index=sysmon to your Sysmon index.

8 more detection rules (Splunk SPL, Microsoft KQL, Sigma) · unlock with TL Creds. Sign up free to unlock the detection rules or query them with the Threadlinqs MCP server (Purple)
  • JavaScript Backdoor Launched by Trojanized AI-Branded Installer (Recipe Lister) — Execution · unlock with TL Creds
  • Rhadamanthys Infostealer [indicator] Outbound C2 from Non-Standard Path — C2 / Credential Theft · unlock with TL Creds
  • Automated Credential Harvesting with Rotating Source IPs Against Cloud Sign-ins (AI Multi-Agent Campaign) — Reconnaissance / Credential Access · unlock with TL Creds
  • Oyster/CleanBoost AutoIt Loader and Fake Dropbox Binary from User-Writable Path — Defense Evasion · unlock with TL Creds
  • AI Assistant Workspace Instruction File Written by Scripting Process (DUSTMAKER-Style Workspace Poisoning) — Supply Chain · unlock with TL Creds
  • Ransom-Note Wallpaper Change by Non-Shell Process (FunkSec) — Impact · unlock with TL Creds
  • FunkSec-Style Ransomware Prep — Defender Tampering Combined with Shadow Copy Deletion — Post-Exploitation · unlock with TL Creds
  • COM Hijacking via InprocServer32 and [indicator] Side-Loading (Oyster / AI-Linked Backdoor Persistence) — Persistence · unlock with TL Creds

Indicators of compromise

As of 2026-11-19, Threadlinqs Intelligence tracks 26 indicators of compromise for TL-2026-6430.

Sample indicators (3 of 26)

  • File hash: 1619bcad3785be31ac2fdee0ab91392d08d9392032246e42673c3cb8964d4cb7
  • File hash: 5226ea8e0f516565ba825a1bbed10020982c16414750237068b602c5b4ac6abd
  • File hash: dcf536edd67a98868759f4e72bcbd1f4404c70048a2a3257e77d8af06cb036ac

Defanged, TLP:CLEAR. Full set of 26 indicators: unlock with TL Creds.

26 indicators of compromise · unlock with TL Creds. Compare plans

Technical analysis

This is a contextual trend-intelligence record, not a single campaign. Halcyon (author Soul Rodriguez, published 2026-10-06) argues that headlines about 'AI malware' suffer from a measurement problem: roughly 97% of AI-connected malware has only ever existed in labs, research repositories, VirusTotal or sandboxes. The 405-sample figure is not named in the Halcyon post, but corresponds to Palo Alto Networks Unit 42's 'The State of AI-Enabled Malware, August 2026' (published 2026-08-25), which analysed 405 unique SHA-256 hashes (WildFire, VirusTotal Intelligence, open-source research) using broad inclusion criteria: AI as a functional component, part of delivery, or part of branding.

Unit 42 found 12 of the 405 samples (3.0%) on Cortex XDR-protected endpoints at organizations in three countries (telemetry Dec 2024 - Jun 2025). The 12 comprised FunkSec ransomware (Rust-based; builds compiled 2025-01-01 to 2025-01-06; disables Windows Defender, deletes volume shadow copies, changes the wallpaper to a ransom note; PDB names Dev.pdb, Funksec.pdb, Darkzone.pdb, Darkfunk.pdb; build cadence suggests LLM-assisted development), a trojanized 'Recipe Lister' AI-branded NSIS installer signed with a since-revoked certificate issued to Global Tech Allies Ltd that runs a JavaScript backdoor (50+ organizations, none executed successfully on protected endpoints), an Oyster/CleanBoost backdoor posing as Dropbox that drops an AutoIt loader and side-loads the backdoor, a Rhadamanthys infostealer (redist.exe), and a COM-hijacking DLL (360Util.dll) impersonating 360 Total Security. Unit 42 states all samples reaching customer environments were blocked, no novel detection approach was needed (sandboxing, behavioral analytics, signer-rarity/code-signing anomalies, entropy analysis), and AI 'changes how the code is authored, not how it executes'. Caveat: the 3% is not an attack success rate and reflects only Palo Alto Networks telemetry.

Halcyon pairs this with three real-world AI-enabled cases. (1) Google Threat Intelligence Group (GTIG, 2026-09-08, Q3 2026 AI Threat Tracker): a suspected financially motivated actor who compromised an organization's cloud environment and deployed an autonomous multi-agent framework driven by an AI coding chatbot and preconfigured markdown 'operational playbooks' to plan, build and run a vulnerability-scanning and credential-harvesting campaign in under six hours, harvesting thousands of third-party credentials with real-time troubleshooting and IP rotation. GTIG separately documents UNC6780/TeamPCP (Altered Spider) supply-chain activity (SANDCLOCK, DUSTMAKER) but has not confirmed a link. (2) Anthropic's August 2025 threat report (2025-08-27): GTG-5004, a UK-based actor with basic coding skills who used Claude to develop and sell ransomware-as-a-service packages (USD 400-1,200, from January 2025) with evasion, encryption and anti-recovery features, and a 'vibe hacking' data-extortion operation (Claude Code automating reconnaissance, credential harvesting and extortion against at least 17 organizations, demands sometimes above USD 500,000; secondary sources attribute it to GTG-2002). (3) An unnamed attacker that used AI to build and launch attacks on print servers across hundreds of organizations (per Halcyon; no primary source identified).

Halcyon Ransomware Research Center data show underground AI-tool advertisements growing from under 50 per month at end of 2025 to over 1,400 by February 2026, largely deepfake video, forged IDs and an 'AI phone bank' (up to 120 simultaneous targets, 25 languages), i.e. identity and social-engineering abuse rather than endpoint malware. Defensive takeaway: AI gives attackers a faster path, not a new one; scouting, initial access, lateral movement, theft and detonation remain observable. Verify requests out-of-band and prioritise behavioral detection and credential/identity hygiene. No CVEs, and no network IOCs suitable for BeaconBeagle correlation, were published in any source.

MITRE ATT&CK techniques by tactic:

Defense Evasion

T1027 Obfuscated Files or Information; T1036 Masquerading; T1574.001 DLL

Execution

T1059.007 JavaScript; T1204.002 Malicious File

Initial Access

T1195.002 Compromise Software Supply Chain

Impact

T1490 Inhibit System Recovery; T1491.001 Internal Defacement; T1657 Financial Theft

Persistence

T1546.015 Component Object Model Hijacking

defense-impairment

T1553.002 Code Signing; T1685 Disable or Modify Tools

Credential Access

T1555 Credentials from Password Stores

Reconnaissance

T1595.002 Vulnerability Scanning

Remediation

Immediate actions

  • Block or alert on the published SHA-256 hashes (FunkSec, Recipe Lister, Oyster/CleanBoost, Rhadamanthys, 360Util.dll)
  • Alert on unsigned/rare-signer installers masquerading as AI applications, Dropbox or security software
  • Verify payment, credential and identity-change requests out-of-band (deepfake/voice-scam defence)

Workarounds

  • Application allow-listing to prevent execution of AI-branded installers from untrusted sources

Longer-term hardening

  • Deploy behavioral/anti-ransomware EDR covering each attack stage rather than relying on static AI-specific signatures
  • Rotate and scope third-party/cloud credentials; monitor cloud accounts for automated scanning and IP-rotation patterns
  • Restrict AI coding agents and CI/CD tokens to least privilege; audit AI assistant workspace configuration files

Affected and fixed versions

  • Various — Windows endpoints, cloud environments and developer/CI-CD credentials targeted by AI-assisted tooling

Timeline

  • FunkSec Rust ransomware variants compiled over 2025-01-01 to 2025-01-06; Unit 42 later cites the build pace as indicative of LLM-assisted development
  • Per Anthropic, the UK-based actor's AI-assisted ransomware-as-a-service offering (USD 400-1,200 per package) first appeared in January 2025 (day approximate)
  • Unit 42 Cortex XDR telemetry window (Dec 2024 - Jun 2025) closes; 12 of 405 AI-linked samples observed on protected endpoints in three countries
  • Anthropic publishes threat report on AI-enabled ransomware development and 'vibe hacking' data extortion of at least 17 organizations
  • Halcyon Ransomware Research Center reports underground AI-tool ads rose from under 50/month (end 2025) to over 1,400 by February 2026 (day approximate)
  • GTIG reports UNC6780/TeamPCP use of SANDCLOCK (CanisterWorm activity) during March-April 2026; DUSTMAKER followed from April 2026 (day approximate)
  • Unit 42 publishes 'The State of AI-Enabled Malware, August 2026' covering 405 samples
  • Google GTIG reports an AI multi-agent framework used to harvest thousands of third-party credentials in under six hours
  • Halcyon publishes 'The AI Malware Panic Has a Measurement Problem', concluding AI offers a faster path, not a new one

Sources cited

Drafted by the Threadlinqs research pipeline from 10 sources; published ; last updated —.

Threadlinqs Intelligence — Real-Time Threat Detection Platform

// threat_feed
$ sort --newest
Showing all threats

Live intelligence console

Threat level

Latest Threats