Google catches AI exploit pipeline running with no human

Published

Google's threat intelligence team says it has, for the first time, caught attackers running an AI framework that finds, weaponizes and fires exploits against websites with no human in the loop.

The finding is the headline of the latest AI Threat Tracker from Google Threat Intelligence Group (GTIG) and Mandiant, written by Aurora Blum, Kelli Vanderlee, Jess Xia and David Mainor. GTIG found the tool, dubbed exploit_agent, on an exposed staging server, alongside data stolen from its victims.

It reads patches to attack the unpatched

exploit_agent is a multi-agent framework wired to several frontier and locally hosted large language models. It is set up to run continuous differential analysis on popular web application software: it compares code from before and after a patch announcement, works out what the patch fixed, drafts an exploit, and executes it automatically.

One example in the report shows the framework inspecting access-control code in version 2.19.25 of the Spectra (Ultimate Addons for Gutenberg) WordPress plugin. The server also held scripts for post-compromise work, including internal reconnaissance, brute-forcing, installing webshells and removing other attackers' webshells.

Victim telemetry on the server suggests the operator was specifically interested in French websites in the media and entertainment industries and in the education sector. Comments in a configuration file holding the LLM API keys were written in Russian and describe the setup as an "exploit_agent pilot". GTIG does not attribute the server to a named threat actor.

The window between patch and exploit keeps shrinking

GTIG presents exploit_agent as one point on a wider curve. Its analysis of vulnerability statistics found that overall disclosures and exploits nearly doubled between January and August 2026. Zero-day exploitation rose more slowly, from a 2025 average of 8 a month to 11 a month in 2026, before spiking to 22 in August.

The report also describes attackers using open source, multi-agent penetration-testing frameworks during live intrusions. GTIG observed CAIRN in intrusion activity against targets including the US aerospace sector, and tracked an actor running a customized PentAGI platform for reconnaissance and vulnerability scanning against organizations in Asia between late August and early October. It has also seen experimentation with CyberStrikeAI and HexStrike AI, and evidence that a PRC-nexus group was building its own automated attack framework. Those tools sit alongside ARTEX, which the FBI warned about this week, and follow earlier reporting on China-linked hackers using CAIRN.

There is a cost for defenders even when the AI output does not work. After two PHP flaws, CVE-2026-109618 and CVE-2026-109619, were patched on October 21, hundreds of claimed but non-functional proofs of concept, likely generated with LLM help, were published, and each one had to be verified or debunked.

Patch faster, and hunt for the noise these tools make

GTIG's assessment is that these frameworks currently trade stealth for velocity. Unrestricted autonomous execution generates forensic traffic, relies on recognizable identifiers and leaves distinct code-level traces, which makes the activity detectable for now. Google says its Security Operations product carries rulepacks for ARTEX, and Mandiant has released a hardening guide aimed at LLM-enabled offensive tooling. GTIG expects the stealth to improve as the underlying models do.

For everyone else the practical answer is speed. If an attacker's pipeline can turn a vendor's patch into a working exploit without a person involved, the days between a plugin update shipping and it being installed are the exposure window. Web-facing CMS plugins, exactly the software exploit_agent was built to read, belong at the front of that queue.

This briefing is provided by IntelFusions for informational and defensive purposes only. It is based on sources assessed to be reliable at the time of writing, and analytic judgments carry the confidence levels indicated. Indicators of compromise are defanged; re-arm them only in controlled environments. IntelFusions is not affiliated with the organizations named and makes no warranty as to completeness or accuracy.

Read the full analysis on IntelFusions