Google: AI agent turned a security patch into a working exploit with no human help
Look ma, no human in the loop!
A security patch inadvertently transformed into a functioning exploit, all without human intervention, according to the Google Threat Intelligence Group (GTIG). This marks the first time such a case has been documented, as revealed in the company's latest AI threat tracker for Q4 2026. The attack employed an AI agent named exploit_agent, built with Russian-language tooling, which continuously analyzed popular web application software by comparing code changes before and after security updates.
This framework identified vulnerabilities, devised exploits, and executed them autonomously. After the attack, the server held stolen data from victim organizations and contained multiple scripts enabling further malicious activities, including reconnaissance, brute-forcing, webshell installation, and webshell removal. The operator's focus on French websites in the media, entertainment, and education sectors raised suspicions.
The report also highlights other instances of attackers utilizing autonomous, multi-agent offensive security frameworks beyond manual LLM prompting, including financially motivated cybercriminals and cyber-espionage groups. Examples of such frameworks include ARTEX, CAIRN, CyberStrikeAI, HexStrikeAI, and PentAGI. GTIG observed attackers using PentAGI in conjunction with proxy services to target the aerospace sector, indicating a growing trend of threat actors experimenting with these frameworks for multi-stage intrusions.
Given the low barrier to entry and ease of integrating new exploits, with vast scaling potential, the Googlers believe that threat actors are highly likely to continue using these frameworks in targeted campaigns globally.
Written by urgent.news from The Register Science's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.