Google Reveals First Fully Autonomous AI Cyberattack: Patches Turned Into Weapons With No Human in the Loop

Google's Threat Intelligence Group reports attackers have used large-language-model frameworks to fully automate vulnerability discovery and exploitation, with no human intervention; researchers call it a "critical inflection point" as defenders' patching windows shrink.
Google Threat Intelligence Group (GTIG) and Mandiant on November 16 released their quarterly AI Threat Tracker report, observing for the first time threat actors using large-language-model frameworks to fully automate vulnerability research and exploitation, with no human in the loop. Researchers described the finding as a "critical inflection point" in the adversarial AI landscape, as machine-speed attacks pull away from human-speed defence.
A tool called exploit_agent that diffs code before and after patches
Between July and September 2026, researchers found a multi-agent framework called exploit_agent on an exposed staging server. Compatible with multiple frontier and local LLMs, it continuously runs differential analysis on patched and unpatched versions of popular web applications, infers the vulnerabilities a patch closes, then drafts and executes exploits automatically. The server also held scripts for post-compromise activity, including internal reconnaissance, brute-forcing and installing webshells. Victim telemetry indicated the operator mainly targeted French websites in the media, entertainment and education sectors; stolen victim data on the server showed the tool had been used against real targets. The code contained Russian-language comments, but Google did not attribute the activity to any organisation or country.
PaperCut flaws were already exploited by fully automated attacks in August
The report cites attacks in August 2026 exploiting vulnerabilities in PaperCut print management servers as evidence that fully automated exploitation is already operational; according to researchers at GreyNoise, the attackers used LLMs to develop exploits and targeted hundreds of organisations in a widespread campaign. The report also notes that zero-day vulnerabilities rose from an average of eight a month in 2025 to 11 in 2026, spiking to 22 in August alone. GTIG rated exploit_agent at Level 5 autonomy and expects more mature fully automated attacks within six months.
Patching windows are shrinking; defence must use AI too
Researchers warned that as automated patch diffing and exploit development tools grow more sophisticated, defenders will have far less time to remediate vulnerabilities; one security chief said traditional patching cycles are becoming obsolete, with some fixes needed in hours rather than days or weeks. The report also notes that today's agentic frameworks trade stealth for speed, generating heavy network traffic and code-level artefacts that make them easier to detect, but that this will change as the tools mature. Researchers urged defenders to bring AI-enabled tools to bear against AI-enabled offence.
(Illustration. Image: Medium.)
N環球 
💬 Comments
No comments yet. Be the first!