ThinkFacility Sign in
  1. Home
  2. News
  3. Google

Google finds attackers using an AI tool that turns security patches into exploits, no human involved

The tool reads each software update, works out which flaw it fixed, and writes the attack itself.

Google's threat intelligence team says it has caught attackers letting AI run the whole job of finding and exploiting software flaws, with nobody in the loop. The AI Threat Tracker report came out on November 16, 2026, from the Google Threat Intelligence Group (GTIG).

GTIG says it's the first time it has seen this. The tool it found reads each software update, works out what the patch fixed, then writes an exploit for the hole and runs it.

Who
Google Threat Intelligence Group
Published
November 16, 2026
Tool
exploit_agent, a multi-agent framework found on an exposed staging server
Targets
French websites in media and entertainment, and the education sector

What Google found

GTIG came across an exposed staging server hosting the framework, which it calls exploit_agent. It plugs into multiple frontier and local models. The report doesn't name which ones.

An aerial view of a cluster of low office buildings with solar panels on their roofs, surrounded by trees, roads and parking, with dry hills behind
Google's main campus in Mountain View, California, from the air. Photo: Austin McKinley, CC BY 3.0, via Wikimedia Commons

The method is patch diffing, done nonstop. The framework compares popular web software before and after a patch is announced, picks out the likely flaw, drafts an exploit and runs it.

GTIG identified threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop for the first time.

From GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold | Google Cloud Blog

It wasn't sitting idle. On the same server, GTIG found data taken from victims and scripts for what attackers do once they're in, from brute-forcing passwords to installing webshells (and deleting other attackers' webshells). One screenshot in the report shows it picking through a WordPress plugin's access checks.

GTIG doesn't name the group behind it or a country. A config file in the report carries Russian-language comments, and that's as far as the report goes.

How fast exploits now arrive

Exploit_agent isn't the only example in the report. In late August 2026, attackers used two flaws in PaperCut print management servers as zero-days, chained for remote code execution. Security researchers and the company Huntress used LLMs to build working exploits within a day of disclosure, and Greynoise reported a campaign against hundreds of organizations four days after it.

GTIG's own count points the same way. It says zero-day exploits averaged 8 a month in 2025 and 11 a month in 2026, then hit 22 in August.

Defending against machine-speed threats at human speed isn’t feasible.

The era of autonomous attacks has arrived.

Who this lands on

The victims GTIG names are ordinary ones: French websites in media and entertainment, and in education. Nothing in the report suggests they were picked for being important. They ran popular web software, and that was enough.

The tool works from patch announcements, so a site that hasn't installed an update is the one it can reach. GTIG's findings say a machine now does the step between a patch notice and a working attack on its own. In the PaperCut case, people with LLMs needed about a day.

The same pressure is behind Anthropic's shorter disclosure deadline, announced the same day. For a school or a small newsroom, I'd read it plainly: install updates the day they come out, or turn on automatic updates.

Google's summary post points defenders to Google's own security products. The full report also covers ARTEX and four other attack frameworks, and AI-run influence operations.

More on Google

All Google stories