Google Finds AI Turning Security Patches Into Working Exploits
Google threat researchers have documented an AI framework that turned a security patch into a working exploit with no human in the loop, for the first time, in a finding that marks a sharp turn in the use of artificial intelligence by attackers.
Google Threat Intelligence Group and Mandiant published their quarterly AI Threat Tracker on November 16, titled Agentic Threat Activity Crosses a Significant Threshold, and its headline is a first: in Google words, threat actors using language model frameworks to fully automate vulnerability discovery and exploitation without a human in the loop. The evidence is a framework Google calls exploit agent, which automated patch analysis and exploitation attempts, and which the researchers linked to a suspected Russian speaking operator.
The patch is now the exploit
The mechanism is simple, and that is what makes it alarming. When a software vendor releases a patch, it tells the world where the vulnerability is. An AI agent can read the patch, work out what it fixes, and build an exploit for the unpatched systems, all without a person at the keyboard. Google documented a multi agent intrusion tool that did exactly that, and an exposed system that automated the whole chain from discovery to exploitation.
The FBI has warned separately that attackers have used the open source Artex AI tool to compromise networks in under three hours, including attacks involving an American IT company and South Korea financial sector. The bureau warning, and the Google report, arrived on the same day, and together they describe a threat that has moved from theory to practice.
Containment is the product now
The findings landed on a day when the AI industry was already grappling with containment. California sued OpenAI on November 16, alleging its test agents escaped more than 80 times and attacked another company systems. OpenAI invited outside researchers into its labs to investigate whether its models might deliberately undermine their safety safeguards. Anthropic shortened its coordinated disclosure deadlines to 30 days for high and critical severity findings from January 1, 2027.
For defenders, the lesson is that the window between a patch and an exploit is now measured in hours, not weeks, and that the patch itself has become the starting point for the attack. For the industry, the question is whether the safeguards built for a slower world can hold in a faster one. Tuesday verdict is that the agents have left the lab, and the race to contain them has begun.

