Google's Threat Intelligence Group (GTIG) says it has discovered an automated AI framework capable of reverse-engineering software security patches and building working exploits entirely on its own.
The tool, named exploit_agent, was found running on an exposed staging server and used multiple large language models (LLMs) to compare code before and after a security update, infer the underlying vulnerability, build an exploit and execute it, all without human intervention.
GTIG said the discovery marks the first time it has observed the complete chain, from patch analysis to live exploitation, fully automated in a single framework, calling it a "significant threshold" in agentic AI's effect on cybersecurity.
The server was first spotted between July and September. It also held stolen data and tools for internal reconnaissance, brute-forcing and webshell deployment, with victims largely drawn from French media, entertainment and education organisations.
Comments embedded in the framework's code were written in Russian, though GTIG has not linked the activity to any specific group or country.
The technique targets so-called n-day flaws, vulnerabilities that vendors have already patched. By reverse-engineering fixes, attackers can build working exploits faster than hunting for unknown zero-days from scratch.
GTIG also flagged other emerging agentic hacking tools, including Artex, Cairn, CyberStrikeAI and PentAGI, built to accelerate multi-stage intrusions.
GTIG draws on Mandiant, Google's Threat Analysis Group, FLARE and VirusTotal, and tracks government-backed hacking, zero-day exploitation and cybercrime affecting Google and its customers.
Read more about Google