Google Threat Intelligence Group confirms: first observation of a fully autonomous vulnerability discovery and exploitation pipeline, says the era of autonomous attacks has arrived
On November 16, 2026, Google Threat Intelligence Group (GTIG), together with Mandiant, released the quarterly report GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold. The report's core finding is that GTIG has, for the first time, observed in the wild threat actors using LLM frameworks to achieve vulnerability discovery and exploitation automation with no human intervention at all.
The specific evidence came from an exposed staging server running a multi-agent vulnerability research and weaponization framework called exploit_agent. The framework connected to multiple frontier and local LLMs and carried out continuous differential analysis on popular web application software: comparing code before and after patches were released, identifying potential vulnerabilities, drafting exploit code, and executing it automatically. The server also contained data stolen from victims, as well as scripts for internal reconnaissance, brute forcing, installing web shells, and removing other attackers' web shells. Telemetry showed that the operator focused on websites related to French media, entertainment, and education.
At the same time, GTIG also confirmed that threat actors have put open-source multi-agent penetration testing frameworks such as ARTEX, CAIRN, CyberStrikeAI, and PentAGI into real-world intrusions, using them to compress multi-stage attack timelines and achieve operations at machine speed. The report says this marks a critical inflection point in the adversarial AI landscape, with attackers shifting from manual prompting to autonomous workflows.
Google's official accompanying blog stated clearly: Defending against machine-speed threats at human speed isn't feasible. It also said that attackers are using autonomous LLM frameworks to continuously scan software updates, identify what patches fix, and build attacks, all without human involvement. GTIG chief analyst John Hultquist added that the main risk at present is the rapid weaponization of n-day vulnerabilities, and that companies must shift to agentic defense.
Source: Google Cloud Blog - GTIG AI Threat Tracker (2026-11-16); accompanying Google Blog statement.
ABAB AI Insight
The report's main incremental value lies in the empirical observation of a closed-loop process with no human involvement, rather than in a mere capability demonstration. Differential analysis itself is a traditional security research technique, but linking it with a multi-agent LLM framework and enabling continuous automated execution compresses the patch window from days or weeks to hours, directly changing the asymmetry between offense and defense. The current framework still trades speed for stealth, leaving obvious traffic and code traces, but the report already predicts that as model capabilities improve, it will both reduce human involvement and increase stealth. For defenders, simply speeding up manual patching can no longer keep pace; Google's AI Threat Defense architecture, promoted in parallel and combining models, context prioritization, automated remediation, and Mandiant intelligence, is a direct response to this.