Google says its analysts believe Russian-speaking hackers carried out the first fully automated AI attack
Google says Russian-speaking hackers ran the first fully automated AI cyberattack
A quarterly report from Google's Threat Intelligence Group and Mandiant says Russian-speaking hackers carried out what its analysts consider the first end-to-end AI cyberattack, with no human involvement throughout the intrusion; the attribution is based on Russian-language comments in the attack code, and the victim data points to French media, entertainment and education websites. The report says the core system, exploit_agent, chains together multiple AI agents, combining frontier models with locally hosted models to continuously compare code before and after patches to commonly used software, and after a patch is released automatically infers the changes, writes an exploit and executes it. Google also observed open-source attack frameworks such as CAIRN and PentAGI being used in campaigns targeting U.S. aerospace organizations and organizations in Asia. The report counts 22 exploited zero-day vulnerabilities in the single month of August, up from 11 per month earlier in 2026 and an average of 8 in 2025; within days after two PHP vulnerabilities were fixed on October 21, hundreds of suspected AI-generated exploit code samples appeared, most of them ineffective.
The report presents the full attack chain from patch diffing to automated exploitation and statistics on zero-day exploitation, which security teams can use to assess the risk of automated attacks within the patch window.