Rank #3 of 64·Security·16 Nov 2026·exploit_agent, an attacker-built multi-agent framework using frontier and local LLMs
Google caught an AI agent turning security patches into working exploits, left on the attackers' own exposed server
“It never sleeps, never eats, and never locked its own server.”
- exploit_agent diffs web-app code before and after each patch
- It drafts exploits and runs them with no human in the loop
- Google calls it the first fully automated pipeline of its kind it has seen
- The exposed server also held stolen victim data
- Scripts for recon, brute-forcing and installing webshells sat alongside it
- Targets: French media, entertainment and education websites
Attackers built an AI agent that reads security patches, works out what they fixed and attacks the hole before anyone updates. Then they left it sitting on an exposed staging server, where Google's threat hunters found it.
What happened
Google Threat Intelligence Group says the framework, which its operator called exploit_agent, plugs into several frontier and local AI models. It keeps watching popular web application software, compares the code before and after each patch announcement, works out which flaw was fixed, drafts an exploit and runs it, with no human in the loop. Google says this is the first time it has seen attackers fully automate both finding and exploiting bugs this way. One example in the report shows the agent picking apart the access checks in a WordPress plugin. A config file carried a Russian comment telling the operator to type the AI keys in right there, "not in the chat".
The damage
This was not a lab demo. The same server held data stolen from victims, plus scripts for internal reconnaissance, brute-forcing, installing webshells and even removing other attackers' webshells. Victim telemetry on the box points to French websites in media, entertainment and education. Google names no victims, no victim count and no group, only a suspected Russian-speaking operator.
How it was caught, and what happened next
Google published the find on 16 November in its quarterly AI Threat Tracker, next to other agent-driven attack frameworks it is following. Its warning for defenders: AI is shrinking the window between a patch going public and an exploit arriving.
Once a patch is public, assume an agent is already diffing it. Patch internet-facing web apps in hours, not weeks, and hunt for webshells.

