█

AI/TLDR

Google Threat Intelligence Group · 2026-11-16 · major

GTIG AI Threat Tracker — attackers now run fully autonomous exploit agents

In short

Google's Threat Intelligence Group found exploit_agent, an attacker framework that reads security patches, writes exploits and runs them with no human in the loop. The report also covers a Claude Code settings.json backdoor.

Google Threat Intelligence Group · 2026-11-16 · 3 sources checked

Google Threat Intelligence banner graphic in red and blue
Google Cloud

Google's quarterly AI threat report says fully automated patch-to-exploit attacks are now happening in the wild.

Quick facts

PublisherGoogle Threat Intelligence Group (GTIG)
PublishedNovember 16, 2026
Headline findingexploit_agent: patch diffing to exploit execution, no human in the loop
exploit_agent targetsFrench media, entertainment and education websites
Frameworks seen in attacksCAIRN, customized PentAGI; tests of CyberStrikeAI and HexStrike AI
Agent backdoorEdited .claude/settings.json to fetch SOMBERMEME

What is it?

The November 2026 GTIG AI Threat Tracker, titled 'Agentic Threat Activity Crosses a Significant Threshold', reports the first case Google has seen of attackers using an LLM framework to fully automate vulnerability discovery and exploitation. Two months earlier, the previous edition said GTIG had not yet seen fully autonomous attack pipelines in the wild.

How does it work?

exploit_agent connects frontier and locally hosted LLMs. It compares popular web applications before and after each patch, finds the likely flaw, drafts an exploit and runs it without a human. Other attackers used open-source agent pentest tools such as CAIRN and a customized PentAGI to automate single phases of an intrusion, trading stealth for speed.

Why does it matter?

The time between a security patch and a working attack is shrinking, so slow patching of internet-facing software is now riskier. Coding-agent config files are also a new place to hide malware: one North Korean cluster planted a backdoor in .claude/settings.json so it ran whenever Claude Code started.

Who is it for?

security teams, platform engineers, teams using coding agents

Frequently asked questions

Who was behind the exploit_agent framework GTIG found?
Google's Threat Intelligence Group did not name a group or country behind exploit_agent. GTIG found it on an exposed staging server, and its configuration file had Russian-language comments telling the operator where to enter LLM API keys. Victim telemetry suggests the operator focused on French media, entertainment and education websites.
How did attackers abuse Claude Code in the GTIG report?
GTIG says a suspected North Korean cluster, MIDNIGHT NEPTUNE, changed .claude/settings.json files inside internal code repositories at cryptocurrency organizations. Each time a developer started Claude Code in that repository, the modified settings downloaded the SOMBERMEME backdoor. The same cluster used fake recruiter profiles whose photos carried Google SynthID watermarks.
Are zero-day or n-day exploits rising faster according to GTIG?
GTIG reports that overall disclosures and exploits nearly doubled between January and August 2026, while zero-day exploits rose more slowly, from a 2025 average of 8 per month to 11 per month in 2026, before a spike to 22 in August. Google suggests LLMs mainly shrink the window between a patch and a working n-day exploit.
How did Google respond to Gemini misuse in this GTIG report?
Google says it disabled the accounts and assets tied to the Gemini API misuse it found, including a contractor's platform for managing more than 100,000 social media accounts and an influence platform aimed at Europe and the Middle East. Google also used the findings to strengthen its classifiers and models.

Try it

Audit .claude/settings.json and other agent hook files in your repos for unexpected commands.

Sources · 3 outlets

Tags

  • google
  • gtig
  • threat-intelligence
  • ai-security
  • agentic-attacks
  • exploitation
  • n-day
  • llm-misuse
  • claude-code
  • supply-chain

← All releases · Learn AI