Google Says It Found an AI Agent Turning Security Patches Into Attacks With No Human Help
Google Threat Intelligence Group says it has seen, for the first time, attackers use an AI framework to find flaws from software patches and exploit them with no human in the loop. The catch for defenders is a patch window that may now close at machine speed.
For years, security people have warned about the day an AI system could read a software patch, work out what bug it fixed, and attack the machines that have not installed it yet, all without a person pressing a button. Google now says that day has arrived. In its latest AI Threat Tracker, published Monday, the Google Threat Intelligence Group said it had seen attackers fully automate that whole chain for the first time.
That is a big claim, and it is Google's claim. GTIG and its Mandiant colleagues are describing what they found, not presenting a court finding. But the wording in the report is blunt.
"In the past quarter, GTIG identified threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop for the first time." Google Threat Intelligence Group
A server someone forgot to lock
The evidence came from an exposed staging server. On it, GTIG said, sat a multi-agent framework the researchers call exploit_agent, hooked up to several frontier and locally run AI models. According to the report, the system watched popular web application software, compared the code before and after patch announcements, drafted exploits for the flaws it spotted, and then ran them automatically with no human intervention.
It was not a lab toy. GTIG said the same server held data taken from targeted victims, plus scripts for follow-on activity once attackers were inside a network. Victim telemetry on the box, the report said, "suggests the operator was specifically interested in targeting French websites associated with the media and entertainment industries, as well as the education sector."
Who ran it is less clear. The Register describes a suspected Russian-speaking operator, based on the Russian-language tooling inside the framework. Google has not tied the activity to any government, and nothing in the report says it was a state operation. Russian comments in a config file are a clue, not a name.
The jump happened fast. In its previous tracker, GTIG said it had documented only "nascent efforts" around AI-assisted bug hunting, including "partially automated vulnerability scanning and exploitation attack sequences." The Register reports that Google framed the new finding as a significant advance on the third quarter. A quarter ago a human still had a hand on the wheel. Now, Google says, nobody did.
There is a cold logic to why attackers would aim AI at patches rather than unknown flaws. GTIG argues that already patched bugs, known as n-days, give a model a finite set of material to chew on, while hunting fresh zero-days means casting a much wider net. The uncomfortable conclusion for every IT team is in the report itself: AI is "shrinking the patch window for organizations." The gap between a fix shipping and someone abusing it used to be weeks. GTIG suggests it may now be however long a model takes to read a diff.
Researchers have shown the raw capability before. Epoch AI said last month that a downloadable Chinese model built working Firefox exploits for about $1,000 each. What GTIG says is new is seeing a hands-off version pointed at real targets in the wild.
The report is not all doom. GTIG said these autonomous tools "currently trade stealth for velocity," leaving forensic traffic and code-level traces that make them easier to spot. That is a thin comfort, and Google admits it may not last.
"As the sophistication of these tools and underlying LLMs increases, we anticipate not only reduced human involvement in these operations but also increased stealth and sophistication of the activity." Google Threat Intelligence Group
The exploit_agent case sits inside a wider pattern GTIG describes, with both criminal crews and suspected espionage groups trying out open source agentic hacking frameworks. Google said it believes it is "highly likely" such frameworks will keep showing up in campaigns worldwide.
For defenders, the takeaway is brutally simple. If Google is right, the patch you meant to install next week is already being read by a machine that does not sleep, does not wait for approval, and does not need you to make a mistake. It only needs you to be slow.

