Google Finds Agents Discovering and Running Exploits With No Human in the Loop
Google Threat Intelligence Group reported on 16 November that it had identified threat actors fully automating vulnerability discovery and exploitation without a human in the loop, for the first time. The evidence was an exposed staging server running a multi-agent framework the group calls exploit_agent, wired to several frontier and local LLMs.
The framework scans popular web application software continuously, compares code before and after patch announcements, works out where the fix implies a flaw, drafts an exploit and executes it. The server also held post-compromise scripts for internal reconnaissance, brute forcing, installing webshells and removing other attackers’ webshells. Victim telemetry points to French websites in media, entertainment and education.
The same report documents open-source offensive frameworks turning up in real intrusions: CAIRN, a multi-agent pentesting loop; CyberStrikeAI; HexStrike AI, an MCP harness fronting more than 150 offensive tools; and PentAGI, which lowers the bar for operators with little experience. GTIG expects use to keep growing.
The surrounding numbers move in the same direction. Vulnerability disclosures and exploits nearly doubled between January and August 2026, and zero-day exploits climbed from a 2025 average of 8 a month to 11 in 2026, then 22 in August. GTIG adds the caveat that automation currently buys speed at the cost of stealth: autonomous execution generates forensic traffic and leaves code-level traces.
Related: Irregular Finds GLM-5.3 Best Open Model on Multi-Stage Cyber Scenarios
GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold · Source channel