Osegi
Security·2 min read

exploit_agent: AI turned security patches into working exploits

Google says a suspected Russian-speaking attacker used an AI agent to read software patches, build exploits and run them, with no human involved, before stealing data.

Lines of code on a dark screen with a red warning icon, illustrating an AI-driven cyberattack.
Image: theregister

Security patches are meant to close holes. Google says one attacker found a way to use them as a map. According to the Google Threat Intelligence Group (GTIG), a suspected Russian-speaking operator used an AI agent called exploit_agent to turn patches into working exploits with no human in the loop. GTIG calls it a first.

Researchers found the setup on an exposed staging server. It hosted a multi-agent framework for finding and weaponizing vulnerabilities, wired up to several frontier and local AI models. It used Russian-language tooling and was set to run continuous differential analysis of "popular web application software". In plain terms, it compared code before and after a security update, spotted the flaw the update fixed, wrote an exploit for it, and then fired it automatically.

The server also held data stolen from victim organizations. There were "multiple" post-compromise scripts too, for internal reconnaissance, brute-forcing, installing webshells (hidden remote-control pages on a web server) and removing other attackers' webshells. Victim telemetry suggests the operator was interested in French websites in media and entertainment, plus the education sector.

Is exploit_agent the only AI attack framework?

No. Google's fourth-quarter 2026 AI threat tracker says attackers are moving beyond typing prompts into a chatbot and are deploying autonomous, multi-agent frameworks for multi-stage intrusions. Both financially motivated criminals and espionage groups are doing it.

Besides the open source penetration-testing tool ARTEX, GTIG lists CAIRN, CyberStrikeAI, HexStrike AI and PentAGI. In one case, attackers used PentAGI with proxy services to target the aerospace sector. Some of the actors experimenting with these tools are suspected espionage groups with a PRC nexus.

What happens next

Google's view is blunt. With a low barrier to entry, easy integration of new exploits and huge potential to scale, it believes threat actors will very likely keep using these frameworks against organizations worldwide. For defenders, the gap between a patch being published and an exploit appearing could get shorter.

Apps worth a look

SUMO: the number puzzle

Tap the right tiles, add them up, hit the target. Easy to learn, hard to master. Free, for all ages.

Tags:googlegtigexploit_agentai agentsvulnerabilitiescyberattack

More from this day

See all →

More in: Security

See all →