Google Finds exploit_agent, an AI That Turns Patches Into Attacks Unaided
Google's Threat Intelligence Group reported on 16 November 2026 that it found exploit_agent, a multi-agent framework on an exposed staging server that compares web application code before and after security patches, drafts exploits, and runs them with no human in the loop. Google calls it the first fully automated discovery and exploitation it has observed. It used both frontier and local models.
On this page
What Google found on the exposed server
The finding comes from the latest quarterly AI Threat Tracker by Google's Threat Intelligence Group (GTIG) and Mandiant, titled "Agentic Threat Activity Crosses a Significant Threshold". In it, GTIG says that in the past quarter it identified threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop "for the first time". Two months earlier, the previous edition said GTIG had not seen fully autonomous attack pipelines used against real targets.
The evidence is a staging server the operator left exposed. It hosted a multi-agent framework called exploit_agent that, in Google's words, "integrates with multiple frontier and local LLMs". The same server held data taken from victims and scripts for internal reconnaissance, brute-forcing, installing webshells, and removing webshells planted by other attackers. Victim telemetry suggests the operator was mainly interested in French websites in media, entertainment, and education.
A configuration file for the model API keys carried comments in Russian. One, as translated by Google, reads: "Enter the keys yourself, right here, not in the chat". Google does not attribute the framework to a named group.
How a patch becomes an attack
The method is patch diffing, an old technique that used to need a skilled person. exploit_agent runs continuous differential analysis of popular web application software: it compares the code before and after a patch announcement, works out what the fix closed, drafts an exploit, and executes it automatically. One figure in the report shows the framework inspecting an access control check in a specific version of Spectra, a widely used WordPress plugin.
GTIG explains why attackers would aim agents at already patched bugs instead of hunting unknown ones. The source material for a patched flaw is finite (two code versions, release notes, public research), while a search for zero-days means casting a much wider net.
The report gives two recent measures of speed. After two PaperCut print server flaws were exploited as zero-days in late August, researchers used LLMs to produce working exploits within a day of disclosure. Four days after disclosure, GreyNoise saw a campaign that likely used LLMs to build exploits and hit hundreds of organizations, with one foothold reaching domain administrator within minutes.
Local models and the limits of account bans
The detail that matters most for the open-weight world is the phrase "frontier and local". When attackers misuse a hosted model, the provider can act: Google says it disabled the accounts and assets tied to the Gemini misuse it found. A model running on the attacker's own hardware has no account to disable.
That is not an argument against open weights, which defenders use for the same analysis. It does mean provider-side enforcement cannot be the main defence, and the report's own framing is about what defenders do on their side. The Decoder notes that Google places current activity at level 5 on its autonomy scale, where tools carry out several attack stages with minimal human involvement, and expects more capable attacks within months.
There is one piece of good news. GTIG says these frameworks currently trade stealth for speed: unrestricted autonomous execution generates forensic traffic and leaves recognizable traces, which makes detection more likely. It expects that advantage to shrink as tools and models improve.
What to do if you run anything on the internet
The lesson is simple and uncomfortable: the day a patch is published is now the day the exploit clock starts. A monthly update routine assumes attackers need weeks. This report describes a system that needs no person at all.
For people who self-host, whether a blog, a home server, or a local AI stack with a web interface, a few habits follow:
- Turn on automatic updates for internet-facing software, and for plugins in particular.
- Remove plugins and services you do not use. Each one is more code for a diffing agent to read.
- Keep admin panels and model servers off the public internet unless they must be reachable; a VPN or private network removes them from mass scanning.
- Be sceptical of proof-of-concept code. GTIG notes that after October's PHP flaws, hundreds of claimed but non-working exploits appeared, likely LLM-generated, and wasted defenders' time.