Google Threat Intelligence Group (GTIG) reported on November 16 that attackers have used an autonomous AI framework, called exploit_agent, to turn security patches into exploits without a human in the loop. It is the first such case GTIG has observed.
The framework continuously compares code before and after a software patch, flags likely vulnerabilities, drafts exploits and runs them. It combines several frontier and local language models. Google found it on an exposed attacker server, where it was aimed at French media, entertainment and education sites.
GTIG calls this a shift from AI-assisted coding to end-to-end automated exploit pipelines. It also notes the tradeoff: fully unrestricted autonomous execution is noisy and leaves recognizable traces, so such tools currently trade stealth for speed. The report also cites hundreds of purported but non-functional proof-of-concept exploits circulating after PHP patches released on October 21.
