---
schema: 1
kind: annual-report
title: "GTIG AI Threat Tracker, November 2026: an exposed exploit_agent framework automates patch diffing, exploit drafting and execution without a human, a suspected North Korean cluster plants a Claude Code startup download in internal repositories, and PRC-nexus actors stage and exfiltrate through an LLM platform's file hosting"
headline: "GTIG quarterly: exploit_agent automates n-day exploitation, suspected DPRK cluster abuses .claude/settings.json"
summary: >
  Google Threat Intelligence Group and Mandiant's quarterly AI threat update of 2026-11-16 reports an exposed multi-agent framework, exploit_agent, that compares code before and after patch announcements, drafts exploits and runs them with no human in the loop, the first such full automation GTIG has identified. It also reports a suspected MIDNIGHT NEPTUNE (North Korea) cluster that modified .claude/settings.json in victims' internal repositories so that Claude Code downloads the SOMBERMEME backdoor at startup, and PRC-nexus actors using an LLM platform's file-hosting backend for staging and exfiltration.
discovered_at: "2026-11-17T04:41:30Z"
updated_at: null
event_date: "2026-11-16"
run_id: 2026-11-17T0355Z-intel
priority: notable
immediate_action: null
tags: [ai-abuse, nation-state, supply-chain]
regions: [global]
sectors: [technology]
entities: ["report:gtig-ai-threat-tracker-nov-2026", "tool:exploit-agent-gtig", "actor:sapphire-sleet", "malware:sombermeme"]
techniques: [T1190, T1587.004, T1505.003, T1566.003, T1204.004, T1195.001, T1098.001, T1567, T1102]
affected_products: ["Anthropic Claude Code"]
cves: []
sources:
  - url: "https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold/"
    publisher: "Google Threat Intelligence Group / Mandiant"
    date: "2026-11-16"
    role: primary
closed_sources: []
evidence:
  - quote: "In the past quarter, GTIG identified threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop for the first time."
    publisher: "Google Threat Intelligence Group / Mandiant"
    source_url: "https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold/"
  - quote: "It compares code before and after patch announcements, analyzes findings to identify potential vulnerabilities, drafts exploits, and automatically executes them without human intervention."
    publisher: "Google Threat Intelligence Group / Mandiant"
    source_url: "https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold/"
  - quote: "so that when users started up Claude Code, it would automatically trigger the download of the SOMBERMEME backdoor"
    publisher: "Google Threat Intelligence Group / Mandiant"
    source_url: "https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold/"
verification: single-source
sourcing_note: >
  GTIG and Mandiant are the first-hand source, drawing on Mandiant incident response and platform telemetry; the report names no victim organisation, and GTIG describes the North Korean cluster as suspected and names no actor for exploit_agent.
confidence: high
references: []
deep_dive: false
deep_dive_category: null
org_triage: null
classification:
  reliability: B
  credibility: 2
watchlist_hit: false
actions: []
updates: []
migrated_from: null
---

Three findings from GTIG's quarterly update on adversary use of AI, published 2026-11-16, change what defenders can look for ([GTIG, 2026-11-16](https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold/)). First, GTIG identified an exposed staging server hosting a multi-agent vulnerability research and weaponization framework, dubbed exploit_agent, that integrates frontier and local LLMs and runs continuous differential analysis of popular web application software: it compares code before and after patch announcements, analyzes the findings, drafts exploits and executes them automatically, which GTIG says is the first time it has seen LLM frameworks fully automate discovery and exploitation without a human in the loop ([GTIG, 2026-11-16](https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold/)). The server also held scripts for internal reconnaissance, brute-forcing, installing web shells and removing other attackers' web shells, and its victim telemetry suggested interest in French media, entertainment and education websites ([GTIG, 2026-11-16](https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold/)).

Second, in intrusions against cryptocurrency organizations in North America and Europe, a suspected MIDNIGHT NEPTUNE cluster used fraudulent recruiter profiles, sent targets to fake meeting platforms with ClickFix lures that ran likely LLM-authored staging scripts, and then pivoted into internal code repositories and modified .claude/settings.json so that Claude Code startup triggered the download of the SOMBERMEME backdoor; it masked lateral movement with GitHub personal access tokens named after Microsoft Copilot ([GTIG, 2026-11-16](https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold/)). Third, in what GTIG calls a new trend, PRC-nexus actors abused LLM platforms and file-hosting backends for staging, dead-drop command channels and exfiltration, in one case staging lure archives and exfiltrating stolen data through a frontier model's file-hosting infrastructure ([GTIG, 2026-11-16](https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold/)). GTIG separately assesses that the open-source agentic pentest frameworks it tracks, among them ARTEX, CAIRN, CyberStrikeAI and PentAGI, currently trade stealth for velocity: they generate forensic traffic, rely on recognizable identifiers and leave distinct code-level traces ([GTIG, 2026-11-16](https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold/)).

**Exposure:** organizations whose developers run Claude Code against shared repositories, where a committed project settings file executes at startup, and internet-facing web applications and plugins whose patches are public, since exploit_agent works from patch diffs.

**Detection:** repository change monitoring for edits to .claude/settings.json and its hook or startup commands, and GitHub audit events that create personal access tokens named after a vendor assistant; for agentic scanning, the forensic traffic, recognizable identifiers and code-level traces that GTIG says pentest frameworks leave behind ([GTIG, 2026-11-16](https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold/)).

**Defender takeaway:** review a change to a project-level Claude Code settings file like a code-execution change, and move internet-facing web components to the fastest patch lane, because GTIG describes patch diffs being turned into executed exploits without a human.
