Cybersecurity researchers at Google LLC's Google Threat Intelligence Group discovered that an artificial intelligence framework turned software security patches into functional exploits without any human intervention. The multi-agent system, identified as exploit_agent, was uncovered on an exposed staging server that also housed data stolen from victim organizations. Google noted that a suspected Russian-speaking operator used the setup to target French websites across the education, media, and entertainment sectors. [1][2]
The framework connected with frontier and local models alongside Russian-language tooling to carry out continuous differential analysis on popular web application code. By comparing software before and after security updates, the tool automatically spotted vulnerabilities, wrote working exploits, and launched attacks. Threat hunters found the staging server configured with post-compromise scripts designed for internal reconnaissance, brute-forcing, installing webshells, and purging competing attackers' webshells. [1][2]
Google documented the incident in its fourth-quarter 2026 AI threat tracker, highlighting a broader shift from manual model prompting to autonomous offensive pipelines. The team noted that both financially motivated criminals and suspected Chinese-nexus espionage groups are testing similar frameworks, including CAIRN, CyberStrikeAI, HexStrike AI, and PentAGI. Google warned that low barriers to entry and massive scaling potential make further global deployments highly likely. [2]