Skip to content

Google finds the first fully autonomous AI exploit pipeline

Google says it found a framework that diffs a security patch, writes a working exploit and runs it without a human, and that monthly zero-day exploitation has nearly tripled this year.

Google’s Threat Intelligence Group says it has found the first fully automated vulnerability research and exploit pipeline running with no human in the loop. The framework, which GTIG calls exploit_agent, compares a patched codebase against the pre-patch version, writes a working exploit for whatever flaw it finds, and runs it, all without a person approving a step. GTIG traced one deployment to a staging server with configuration notes in Russian, targeting French media, entertainment and education sites.

Published on 17 November, the report also gives numbers on where exploitation generally is heading: zero-day exploits averaged eight a month in 2025, rose to eleven a month so far in 2026, then spiked to 22 in August. Vulnerability disclosures and exploits nearly doubled between January and August. Google points to the PaperCut vulnerability chain as the clearest case: two flaws were being exploited as zero-days by late August, researchers had working exploits within 24 hours, and in one attack a foothold became domain admin access within minutes. GTIG also names two open-source multi-agent pentesting frameworks, CAIRN and PentAGI, as already deployed in real intrusions rather than just tested.

Why it matters: the gap between a patch landing and someone weaponising it is now measured in hours, not weeks. If you run internet-facing software, “patch at the next maintenance window” is no longer a safe default, especially once a proof-of-concept exploit is public, AI-written or not.

The caveat: Google also sells the Gemini-based defences, Big Sleep and CodeMender, that this same report recommends. That doesn’t make the underlying numbers wrong, but it’s worth remembering who is doing the counting.

Share

Related reading

More from The Wire

All briefs