Attackers built an AI framework designed to turn software patches into exploits automatically, Google's Threat Intelligence Group disclosed Monday. Google says this is the first time it has identified attackers using LLM frameworks to fully automate vulnerability discovery and exploitation "without a human in the loop." Researchers discovered it on an exposed staging server alongside stolen victim data and post-compromise scripts, but the report does not establish that the framework itself was responsible for those compromises.

The framework, identified as exploit_agent in its configuration, integrates multiple frontier and local LLMs, according to Google. It is set up to continuously compare code before and after software patches, identify potential vulnerabilities, draft exploits and execute them without human intervention.

Google did not identify a specific vulnerability exploited by the framework, provide a victim count or report a measured patch-to-exploit time.

"Attackers didn't invent a new kind of bug. They automated the tedious part: reading the patch, finding what it fixed, and writing the exploit," said Merritt Baer, former Deputy CISO at AWS and executive advisor to Upwind and G2i. "A skilled person might spend days on that. Now an agent does it while its operator sleeps."

What Google actually found

Google found that the operator's configuration file contains Russian-language comments describing exploit_agent as a pilot and instructing the operator to enter LLM API keys directly into the file. Google does not attribute the framework to a particular group or country. Victim telemetry on the server suggests the operator was interested in targeting French media, entertainment and education websites, according to Google.

Google also published a code excerpt showing the framework examining access controls in version 2.19.25 of the Spectra WordPress plugin, but did not tie it to a successful exploit.

Google's previous threat intelligence reports documented partially automated vulnerability scanning and exploitation. Its latest report also describes AI frameworks, including CAIRN and PentAGI, being used in live intrusions. What distinguishes exploit_agent is its specific design to turn software patches into exploits.

"None of the response is exotic. Know what you expose to the internet, patch it faster than you did last year, and watch for exploitation, not just for vulnerabilities," Baer said. "Attackers bring speed; defenders have to bring knowledge of their own environment."