# GTIG warns fully-automated agentic AI attacks imminent: exploit_agent, CAIRN, ARTEX abuse and SOMBERMEME via modified Claude Code (MIDNIGHT NEPTUNE / UNC6836)

> Google Threat Intelligence Group's AI Threat Tracker (2026-11-16) reports that threat actors fully automated vulnerability discovery and exploitation without a human in the loop for the first time, and estimates low-sophistication attacks needing little human interaction within ~3 months and sophisticated fully autonomous attacks within ~6 months. Observed activity includes the exploit_agent framework, abuse of CAIRN/ARTEX/PentAGI/CyberStrikeAI/HexStrike AI, and a suspected DPRK cluster (MIDNIGHT NEPTUNE, reported by BleepingComputer as UNC6836) delivering the SOMBERMEME backdoor via modified Claude Code settings.

- **Published:** 2026-11-16T14:40:18Z
- **Last reviewed:** 2026-11-17T05:57:26.192Z
- **Canonical:** https://intel.threadlinqs.com/threat/TL-2026-6152
- **ID:** TL-2026-6152
- **Severity:** HIGH
- **Category:** THREAT_INTEL
- **Status:** ACTIVE
- **Actor:** MIDNIGHT NEPTUNE (North Korea)
- **Detections:** 9 (1 shown free below) · **IOCs:** 59 (3 shown free below)
- **CVEs:** CVE-2026-81578, CVE-2026-82078, CVE-2026-109618, CVE-2026-109619, CVE-2026-40217, CVE-2021-42278, CVE-2021-42287

## Description

GTIG's AI Threat Tracker, 'Agentic Threat Activity Crosses a Significant Threshold' (Blum, Vanderlee, Xia, Mainor; 2026-11-16), documents a shift from AI-assisted to agent-driven intrusion activity. GTIG states that in the past quarter it observed threat actors fully automate vulnerability discovery and exploitation without a human in the loop for the first time, and that these frameworks sharply shorten the time defenders have to patch.

exploit_agent is a multi-agent vulnerability research and weaponization framework found on an attacker's exposed staging server. It integrates with frontier and local LLMs, performs continuous differential analysis of web application updates, drafts exploits and executes them without human intervention. The staging server held reconnaissance output, stolen victim data, brute-forcing scripts and webshell installation scripts, including scripts that remove other attackers' webshells. Victims were French media, entertainment and education websites. GTIG also describes an agentic framework analyzing Spectra (Ultimate Addons for Gutenberg) v2.19.25 (file v21925.php), a WordPress plugin with an insecure access-control check.

Open-source offensive frameworks are being abused: CAIRN (multi-agent pentest framework using an OODA loop and shared blackboard; observed in intrusions against the U.S. aerospace sector), ARTEX (Go-based autonomous pentest framework weaponized by multiple actors since August 2026; an FBI FLASH advisory was released 2026-11-13 and Google SecOps has deployed detection rulepacks), PentAGI (a customized version used for recon and scanning of organizations in Asia from late August to early October 2026), CyberStrikeAI (LLM-native framework with a Go backend acting as C2/orchestration) and HexStrike AI (MCP-integrated harness connecting LLMs to 150+ offensive tools via 12+ sub-agents), the latter two seen in experimentation. Per the FBI advisory as reported by BleepingComputer, ARTEX agents perform recon, exploitation, privilege escalation and data theft, including password spraying, cloud application permission abuse, credential modification and web shell deployment, reaching privileged access in under three hours. Separate Gambit reporting describes a financially motivated operator chaining Strix (scanning), Cairn (exploitation) and Hermes (orchestration) to compromise at least 27 companies and inject skimmers into 119+ sites.

GTIG attributes SOMBERMEME delivery to MIDNIGHT NEPTUNE, a suspected DPRK cluster that combines LLM-enabled social engineering with abuse of AI tools as a supply-chain vector: the actor modified .claude/settings.json in internal repositories so that launching Claude Code triggers download of the SOMBERMEME backdoor. Targets are cryptocurrency organizations in North America and Europe. BleepingComputer reports the cluster as UNC6836 with a suspected link to Midnight Neptune; the GTIG article itself does not use the UNC6836 identifier. GTIG additionally reports LLMjacking (theft of LLM credentials and cloud resources resold via LLM Mixers) with single-day charges over $100,000 and cumulative losses up to $1M, and LLM-assisted exploitation of PHP flaws (CVE-2026-109618/-109619; patches 2026-10-21) and PaperCut zero-days (CVE-2026-81578, CVE-2026-82078) exploited four days after disclosure. No hashes, IPs or domains are published in the sources reviewed; IOCs here are tool, malware, filename and entity indicators. Attribution of SOMBERMEME to DPRK is 'suspected'.

## MITRE ATT&CK

- T1595 Active Scanning
- T1587.004 Develop Capabilities: Exploits
- T1190 Exploit Public-Facing Application
- T1195 Supply Chain Compromise
- T1566 Phishing
- T1078.004 Valid Accounts: Cloud Accounts
- T1204 User Execution
- T1505.003 Server Software Component: Web Shell
- T1136.003 Create Account: Cloud Account
- T1098 Account Manipulation
- T1036 Masquerading
- T1110.003 Brute Force: Password Spraying
- T1552 Unsecured Credentials
- T1090 Proxy
- T1496 Resource Hijacking
- T1595.002 Active Scanning: Vulnerability Scanning
- T1588.007 Obtain Capabilities: Artificial Intelligence
- T1583 Acquire Infrastructure
- T1650 Acquire Access
- T1068 Exploitation for Privilege Escalation
- T1539 Steal Web Session Cookie
- T1552.001 Unsecured Credentials: Credentials In Files
- T1567 Exfiltration Over Web Service
- T1566.003 Phishing: Spearphishing via Service
- T1195.002 Supply Chain Compromise: Compromise Software Supply Chain
- T1204.004 User Execution: Malicious Copy and Paste
- T1059 Command and Scripting Interpreter
- T1562.008 Impair Defenses: Disable or Modify Cloud Logs
- T1036.005 Masquerading: Match Legitimate Resource Name or Location
- T1528 Steal Application Access Token
- T1558.003 Steal or Forge Kerberos Tickets: Kerberoasting
- T1003.006 OS Credential Dumping: DCSync
- T1550.002 Use Alternate Authentication Material: Pass the Hash
- T1071.001 Application Layer Protocol: Web Protocols
- T1572 Protocol Tunneling
- T1105 Ingress Tool Transfer
- T1059.001 Command and Scripting Interpreter: PowerShell
- T1098.001 Account Manipulation: Additional Cloud Credentials
- T1003.004 OS Credential Dumping: LSA Secrets
- T1649 Steal or Forge Authentication Certificates

## Free detection (Splunk SPL, Microsoft KQL, Sigma)

**SOMBERMEME Delivery - Claude Code Launch Spawning Download Child Process (TL-2026-6152)** (`TL-2026-6152-SPL-001`)

Detects the supply-chain delivery pattern attributed to MIDNIGHT NEPTUNE (suspected DPRK) where a modified .claude/settings.json in an internal repository causes Claude Code startup to download the SOMBERMEME backdoor. The search looks for process-creation events where claude (or node running claude) is the parent and the child command line contains download primitives (URL, DownloadString, DownloadFile, Invoke-WebRequest, certutil urlcache, bitsadmin). Behaviour-first: it does not depend on any published hash or domain, because GTIG published none.

- **Languages:** Splunk SPL, Microsoft KQL, Sigma
- **ATT&CK:** T1195, T1204, T1036
- **Data sources:** sysmon, microsoft_defender_for_endpoint
- **Severity:** high
- **Confidence:** high

### Splunk SPL

```spl
index=sysmon sourcetype="XmlWinEventLog:Microsoft-Windows-Sysmon/Operational" EventCode=1 earliest=-24h (ParentImage="*claude.exe" OR ParentImage="*/claude" OR (ParentImage="*node*" AND ParentCommandLine="*claude*")) (CommandLine="*http://*" OR CommandLine="*https://*" OR CommandLine="*DownloadString*" OR CommandLine="*DownloadFile*" OR CommandLine="*Invoke-WebRequest*" OR CommandLine="*urlcache*" OR CommandLine="*bitsadmin*") | stats count min(_time) as first_seen max(_time) as last_seen values(CommandLine) as child_commands values(ParentCommandLine) as claude_cmdline by host, User, ParentImage, Image | eval first_seen=strftime(first_seen,"%Y-%m-%d %H:%M:%S"), last_seen=strftime(last_seen,"%Y-%m-%d %H:%M:%S") | table first_seen last_seen host User ParentImage Image child_commands claude_cmdline count
```

### Microsoft KQL

```kql
DeviceProcessEvents
| where Timestamp > ago(24h)
| where InitiatingProcessFileName in~ ("claude.exe", "claude") or (InitiatingProcessFileName in~ ("node.exe", "node") and InitiatingProcessCommandLine has "claude")
| where FileName in~ ("curl.exe", "curl", "wget", "powershell.exe", "pwsh.exe", "pwsh", "cmd.exe", "bash", "sh", "certutil.exe", "bitsadmin.exe", "mshta.exe")
| where ProcessCommandLine matches regex @"(?i)(https?://|DownloadString|DownloadFile|Invoke-WebRequest|\biwr\b|\birm\b|urlcache|bitsadmin|curl |wget )"
| project Timestamp, DeviceName, AccountName, InitiatingProcessFileName, InitiatingProcessCommandLine, FileName, ProcessCommandLine, SHA256
```

### Sigma

```yaml
title: Claude Code Spawning Download Child Process - SOMBERMEME Delivery Pattern
id: 7c1e4a52-3b8d-4f6a-9e21-5d0a8c3b7f14
status: experimental
description: Detects Claude Code (or node running Claude Code) launching a child process that downloads content, the delivery pattern GTIG attributes to MIDNIGHT NEPTUNE abusing a modified .claude/settings.json to fetch the SOMBERMEME backdoor.
references:
  - https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold
  - https://www.bleepingcomputer.com/news/security/google-warns-that-fully-automated-ai-attacks-are-coming-soon/
author: Threadlinqs Team
date: 2026-11-16
tags:
  - attack.initial-access
  - attack.t1195
  - attack.execution
  - attack.t1204
  - attack.defense-evasion
  - attack.t1036
logsource:
  category: process_creation
  product: windows
detection:
  selection_parent_claude:
    ParentImage|endswith: '\claude.exe'
  selection_parent_node:
    ParentImage|endswith: '\node.exe'
    ParentCommandLine|contains: 'claude'
  selection_download:
    CommandLine|contains:
      - 'http://'
      - 'https://'
      - 'DownloadString'
      - 'DownloadFile'
      - 'Invoke-WebRequest'
      - 'urlcache'
      - 'bitsadmin'
  condition: (selection_parent_claude or selection_parent_node) and selection_download
falsepositives:
  - Developers asking Claude Code to run curl or package installs that fetch from URLs
  - Legitimate Claude Code hooks that call internal tooling over HTTP
level: high
```

**Known false positives:**

- Developers legitimately instructing Claude Code to run curl, wget or package managers that fetch from URLs
- Sanctioned Claude Code hooks in .claude/settings.json that call internal tooling over HTTP

**Tuning notes:** Baseline approved hook commands per repository and allow-list known internal download hosts. Pair with SIG-001 (settings.json modification) to correlate the config change with the first launch. Extend ParentImage patterns if Claude Code is installed under a different binary name on your fleet.

### 8 more detection rules (Splunk SPL, Microsoft KQL, Sigma), unlock with TL Creds

- Agentic Web Exploitation - High-Velocity Path Probing and Spectra [indicator] Access (locked: unlock with TL Creds)
- Password Spraying Against Entra ID - Many Accounts Failing From One IP (locked: unlock with TL Creds)
- Entra ID Cloud Application Permission Abuse and Credential Modification Burst (locked: unlock with TL Creds)
- Claude Code settings.json Written by Script Interpreter or Download Utility (SOMBERMEME Supply-Chain Vector) (locked: unlock with TL Creds)
- AI Offensive Framework Execution - ARTEX, CAIRN, PentAGI, CyberStrikeAI, HexStrike AI, exploit_agent (locked: unlock with TL Creds)
- Community-Observed DPRK-Tagged Infrastructure and File Hash Watchlist (Low Confidence) (locked: unlock with TL Creds)
- Web Shell Activity - Web Server Process Spawning Shell or Recon Utilities (exploit_agent / ARTEX) (locked: unlock with TL Creds)
- LLMjacking - AWS Bedrock Model Enumeration, Burst Invocation or Invocation Logging Tampering (locked: unlock with TL Creds)

Sign up free to unlock the detection rules: https://threadlinqs.com/?modal=signup

## Sample indicators (3 of 59)

Defanged, TLP:CLEAR.

- IP address: `45.142.193[.]132`
- Domain: `jet-night[.]com`
- File hash: `528cd4e69ecfa5191adbcf6ef28667bf`

Full set of 59 indicators unlocks with TL Creds: https://threadlinqs.com/tl-creds

## Sources

- [Google warns that fully-automated AI attacks are coming soon (BleepingComputer)](https://www.bleepingcomputer.com/news/security/google-warns-that-fully-automated-ai-attacks-are-coming-soon/)
- [GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold](https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold)
- [FBI warns ARTEX AI hacking tool is powering global cyberattacks (BleepingComputer)](https://www.bleepingcomputer.com/news/security/fbi-warns-artex-ai-hacking-tool-is-powering-global-cyberattacks/)
- [Malicious AI agents steal 600K credit cards, infect 100+ sites with skimmers (BleepingComputer)](https://www.bleepingcomputer.com/news/security/malicious-ai-agents-steal-600k-credit-cards-infect-100-plus-sites-with-skimmers/)
- [Google: LLMjacking attacks rack up $1 million in cloud AI charges (BleepingComputer)](https://www.bleepingcomputer.com/news/security/google-llmjacking-attacks-rack-up-1-million-in-cloud-ai-charges/)

## Full data

The other 8 detection rules (Splunk SPL / Microsoft KQL / Sigma) unlock with TL Creds; a free account includes daily creds: https://threadlinqs.com/?modal=signup
Indicator values unlock with TL Creds, at a lower price on Red and Purple; AI agents can query rules and IOCs through the Threadlinqs MCP server (Purple): https://intel.threadlinqs.com/mcp

Canonical: https://intel.threadlinqs.com/threat/TL-2026-6152
