LONDON - Google Threat Intelligence Group and Mandiant reported on Nov. 16 that hackers deployed a custom multi-agent AI system called exploit_agent. The tool automated vulnerability discovery, exploit execution and data theft. No human directed the process.
Exposed Staging Server Reveals Automated Patch-to-Exploit Pipeline
Investigators found exploit_agent on an exposed attacker staging server. The system chains frontier and local language models. It compares code before and after vendor security patches. It then launches attacks against internet-facing web applications.
The server also held stolen victim data, brute-forcing scripts and webshells. It carried a tool to remove rival attackers' webshells. Google withheld affected product names and vulnerability identifiers.
"As of 16 November 2026, the company assessed it as the first observed operation to conduct fully automated vulnerability discovery and exploitation without a person directing the process," CyPro reported. The outlet summarized Google's findings.
Russian-Language Clues Target French Sectors as Patch Windows Shrink
Code comments and instructions inside the toolkit contained Russian-language text. Recovered artifacts showed interest in French media, entertainment and education websites. Google did not attribute the campaign to a specific group or state.
Google also noted that hundreds of AI-generated proof-of-concept exploits appeared within days. They surfaced after two PHP flaws were patched on Oct. 21. The flaws are CVE-2026-109618 and CVE-2026-109619.


