THE CRUNCH

Google's Threat Intelligence Group (GTIG) says it has documented what it describes as the first case of an AI agent turning security patches into working exploits with no human in the loop, before data was stolen from victim organisations. The finding appears in the company's fourth-quarter 2026 AI threat tracker, published on Monday, 16 November 2026.

The trail began with an exposed staging server hosting a multi-agent vulnerability research and weaponisation framework, which the report names exploit_agent. It integrates with multiple frontier and local models, used Russian-language tooling, and was configured to run continuous differential analysis of popular web application software: comparing code before and after security updates, identifying potential vulnerabilities, developing exploits and then executing them automatically.

The server also held data stolen from victims and multiple post-compromise scripts for internal reconnaissance, brute-forcing, installing webshells and removing other attackers' webshells. Victim telemetry on the server suggested the operator was specifically interested in French websites tied to media, entertainment and education.

Google framed the case as a significant advance on the third quarter in how attackers use large language models for fully autonomous vulnerability discovery and exploitation. The same report documents other groups moving beyond manual LLM prompting to autonomous multi-agent offensive frameworks, including CAIRN, CyberStrikeAI, HexStrike AI and PentAGI, used by both financially motivated criminals and suspected PRC-linked espionage actors. In one instance, attackers paired PentAGI with proxy services to target the aerospace sector.