Radar· local AI & agent security
security high confidence

Google: first documented pipeline that turns a patch into an exploit and attacks with no human in the loop

Nov 16, 2026 · agents, attack, gtig

TL;DR A Google Threat Intelligence/Mandiant report describes agents that diff patches, build the exploit and run it on their own. The same report covers a tampered .claude/settings.json.

In plain words: when a vendor ships a security fix, the fix itself shows where the hole was. Attackers now use AI agents to read the fix, write the attack and launch it automatically, so the time you have to update keeps shrinking.

What happened

A report from Google Threat Intelligence and Mandiant describes:

  • a pipeline that finds the flaw by comparing patches, builds the exploit and runs it, with no human in between, using multi-agent frameworks (ARTEX, CAIRN, CyberStrikeAI, PentAGI);
  • a group suspected to be North Korean that changed the .claude/settings.json of internal repositories at crypto companies so that every Claude Code session downloaded a backdoor;
  • Microsoft reporting AI-assisted intrusions reaching privileged access in under 3 hours.

Who is affected

Everyone who patches on a weekly or monthly rhythm, and every team whose repositories carry agent configuration files.

Why it matters

The gap between a published patch and a working exploit is now measured in hours. And agent configuration inside a repository is executable in practice: whoever edits it decides what the agent runs.

What to do

  1. Shorten the patch cycle for internet-facing systems.
  2. Review agent configuration files (.claude/settings.json, hooks, MCP configs) in code review like any other code.

Sources

← back