Google reports AI agent running exploits with no human involved
Google and Mandiant describe the first documented end-to-end attack carried out by an AI agent, with no human involved at each step.
The report from the Google Threat Intelligence Group, published on 16 November, describes a multi-agent system called exploit_agent, found on an exposed staging server. It compares code before and after a patch, writes exploits and runs them on its own. Targets are mainly French media, entertainment and education sites. Google names no group or state, despite Russian-language comments in the code. Help Net Security reported it on 17 November.
Source: Help Net Security, published on November 17, 2026.