Google says Russian-speaking hackers ran first fully automated AI cyberattack
For defenders, the ugly part is speed: patch-diffing is becoming an automated exploitation queue before triage catches up.
TL;DR
CyberScoop reports that Google’s Threat Intelligence Group and Mandiant found what Google calls the first end-to-end AI-run intrusion, tied to Russian-language code but not to a named group or state. Victim data pointed to French websites in several civilian sectors; Google also observed CAIRN and PentAGI in activity aimed at a U.S. aerospace organization and organizations across Asia. The uncomfortable bit isn’t attribution. It’s public patching becoming a machine-readable exploit cue.
Google's finding matters because it moves AI misuse from assistance to orchestration, at least in Google's telling. CyberScoop reports that the company has not named a group or state, which is the right caveat for compliance teams that have seen too many inflated AI-attack claims. The evidence Google described is narrower and more useful: Russian-language comments and an auto-exploitation configuration file; victim data pointed to French websites in media, entertainment and education.
The exploit_agent system Google describes compares code before and after a vendor patch, infers what changed, drafts an exploit and runs it. That sentence is boring if you're a patch manager, which is the problem. Patch publication has always been a race. The report says the other side is turning the starting gun into part of the machine.
Google put numbers around the broader acceleration: 22 exploited zero-days in August, compared with 11 per month earlier in 2026 and an average of eight in 2025. It also saw open-source offensive frameworks including CAIRN and PentAGI in activity targeting a U.S. aerospace organization and organizations across Asia. None of that proves every AI-assisted intrusion is clever. Loud still works if it gets there first.
For federal contractors and agencies consuming the same software, the Monday task is basic and unpleasant. Treat public patches as adversary inputs. Shorten exposure windows for internet-facing software, and make sure incident responders can tell noisy automation from background scan trash before the alert queue becomes an archaeology project.
Published ·Deep Fathom