Google Finds AI Framework That Turns Patches Into Exploits, No Human Needed
Google says attackers built exploit_agent, an AI framework that reads security patches, spots flaws and fires off working exploits without a human in the loop. Researchers found it on an exposed staging server sitting next to stolen victim data. Google calls it the first confirmed case of fully automated n-day exploitation.
- 2 sources report it: thedefiant.io, X account @Cointelegraph.
- The source published it on Nov 16, 2026 at 2:08 PM.
Nothing open. Every check passed.
Google Threat Intelligence Group dropped the details in a Nov. 16 report. The team says it found exploit_agent on an exposed staging server, stashed alongside data stolen from victims. That same server carried scripts for surveying compromised networks, brute-forcing access, installing webshells and deleting other attackers' webshells.
According to Google, the framework watches web application code before and after patch announcements, hunts for flaws in the changes, drafts exploits and runs them automatically. The target is already-disclosed “n-day” vulnerabilities, not unknown bugs. Google's evidence shows it inspecting access controls in version 2.19.25 of Spectra, a WordPress plugin also called Ultimate Addons for Gutenberg, plus configuration-file comments for multiple model API keys. The framework links to multiple frontier and locally hosted language models, which Google does not name.
Victim telemetry pointed to interest in French media, entertainment and education websites, Google said. But the company does not tie exploit_agent to a specific group, name a CVE for the activity or connect any generated exploit to a particular victim. GTIG called this the first time it had seen threat actors fully automate vulnerability discovery and exploitation “without a human in the loop.” Google says the finding extends its earlier reports on AI-assisted vulnerability research and agents handling parts of live intrusions, not the first use of AI in an attack.
The defensive headache: a published fix gives attackers a bounded research task. Google argues that comparing code versions, release notes and related research may be more efficient for language models than searching broadly for unknown flaws. It did not give a measured patch-to-exploit time for exploit_agent.
A separate PaperCut campaign shows the pace. GreyNoise reported an actor using AI agents to compromise systems at 395 identified organizations, with exploit development starting four days after the print-software vendor's advisory. GreyNoise's timeline includes a prompt asking the human operator for permission to continue, a contrast with Google's account of exploit_agent. Mandiant's defensive guidance calls for short-lived credentials, restricted privileges, network segmentation and behavioral detection, describing them as long-standing practices rather than new controls for AI attacks.
TopicsGoogle · Google Threat Intelligence Group · exploit_agent · Spectra · Ultimate Addons for Gutenberg · GreyNoise · Mandiant · PaperCut
Written by AI from verified reporting.

