GTIG says it first found an attacker using exploit_agent in a pilot to fully automate patch-to-exploitation
exploit_agent: Autonomous Patch-to-Exploit Pipelines Arrive
Google Threat Intelligence Group (GTIG) and Mandiant reported in the AI Threat Tracker released on November 16, 2026, that they had for the first time found a threat actor using an LLM framework to fully automate vulnerability discovery and exploitation without human involvement. The framework, exploit_agent, was recovered from an exposed staging server; it continuously compares code from popular web application software before and after patch releases, drafts exploits, and executes them automatically, and according to victim telemetry its operators focused on websites in the French media, entertainment, and education sectors. GTIG did not disclose success rates or attribution, said such tools currently trade stealth for speed, and listed them alongside agentic attack frameworks such as CAIRN, PentAGI, ARTEX, and HexStrike AI. CSA recommends treating patch releases for internet-facing software as imminent exploitation events, shortening response to the hour level, and auditing AI API credentials to prevent LLMjacking; GTIG recorded such incidents with single-day fraud costs exceeding 100,000 US dollars.
GTIG's findings indicate that patch releases can serve as machine-readable signals for automated attacks, and CSA accordingly proposes exposure-driven patch response measured in hours and detection recommendations for noisy behavior.