Google's Threat Intelligence Group has documented the first fully automated AI cyberattack. A system called exploit_agent, attributed to Russian-speaking threat actors, combines frontier and locally hosted language models into an attack pipeline that operates without any human in the loop. It compares pre- and post-patch code to identify exploitable deltas, drafts working exploits, and executes them, all automatically.
How exploit_agent works
The system chains multiple AI agents into a sequential pipeline. One agent monitors for newly disclosed vulnerabilities and pulls patch diffs. A second analyzes the differences to find exploitable code paths. A third drafts exploit code tailored to the target. A fourth tests and executes. The result is an end-to-end attack chain that runs faster than any human analyst can track, operating continuously from disclosure to exploitation.
The August 2026 campaign
According to the Google TIG report, exploit_agent was used to successfully exploit 22 zero-day vulnerabilities in August 2026. Targets were concentrated in French media, entertainment, and education organizations. In each case, the system operated independently, selecting and executing exploits without human review.
What this means for your patch window
Historically, the race between patch release and exploitation gave defenders days or weeks to remediate. That window is now hours at most. Exploit_agent compresses the attack cycle to the speed of automated code analysis. If a critical patch ships on Tuesday morning, an AI attacker can have a working exploit by Tuesday afternoon. Organizations that run weekly or monthly patch cycles are structurally exposed.
What defenders should do
Treat CISA KEV items as requiring a 24-hour remediation cycle, not a weekly one. Prioritize externally facing systems and authentication endpoints above all else. Where patching is not immediately possible, compensating controls such as WAF rules, rate limiting, and anomaly detection become load-bearing. Maintain accurate, current asset inventory: you cannot patch what you cannot see.
Gigia Tsiklauri is a Security Architect and founder of Infosec.ge. Get in touch if you want to discuss how to compress your patch cycle time and build an automated vulnerability response workflow.