# Google Finds AI Framework Configured to Automate Patch-Based Exploits.

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/59718
Published: 2026-11-17T18:07:54.000Z
Updated: 2026-11-17T18:07:54.000Z
Section: Technology

> The setup linked frontier and local LLMs in continuous reviews of web application code around patch announcements. The server also held victim data and scripts for reconnaissance, password attacks and webshell operati…

Google Threat Intelligence Group (GTIG) and Mandiant found an exposed staging server hosting a multi-agent AI framework configured to identify vulnerabilities through patch analysis and execute exploits automatically. GTIG identified this as the first case it had seen of threat actors using LLM frameworks to automate vulnerability discovery and exploitation without human intervention.

The framework, named exploit_agent, connected to frontier and locally run large language models. Its configuration described continuous differential analysis of popular web application software, examining code from before and after patch announcements. The process was set up to identify potential weaknesses in changes around fixes, then draft exploit code and run it.

Those details describe the framework’s configuration, but the findings do not provide an exploit-run count, a success rate or a specific vulnerability successfully exploited by exploit_agent. The findings do not quantify whether the framework successfully exploited any vulnerabilities.

The server also contained scripts for internal reconnaissance and password brute-forcing. Some scripts could install webshells, while others could remove webshells left by other attackers. GTIG identified data on the server as exfiltrated from targeted victims, but did not attribute the theft to exploit_agent or provide a count of affected organizations.

GTIG said server telemetry suggested the operator was interested in targeting French websites in media, entertainment and education. The signal indicates interest in those sectors; it does not establish that every site there was compromised.

Vulnerability disclosures and exploits nearly doubled between January and August 2026. Zero-day exploits averaged 11 per month in 2026, compared with eight per month in 2025, and reached 22 in August. These figures describe broader vulnerability activity, not the number or success of runs by exploit_agent.


