Illustration of an autonomous AI agent in a dark server room comparing two versions of software code and building a glowing exploit, representing Google GTIG's exploit_agent finding

The patch is supposed to keep you safe. Google just found an AI that reads the patch, works out what it fixed — and builds the attack itself, with nobody at the keyboard.

In a report published November 16, Google Threat Intelligence Group (GTIG) and Mandiant say: “In the past quarter, GTIG identified threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop for the first time.” Two months ago, GTIG’s previous report said it had not seen fully autonomous attack pipelines in the wild. That line has now been crossed.

How exploit_agent works

GTIG found the framework, dubbed exploit_agent, on an exposed staging server. It is wired into multiple frontier and locally hosted LLMs and is set up for continuous differential analysis of popular web application software: compare code before and after a patch, spot the likely vulnerability, draft an exploit, and automatically execute it — without human intervention.

On the same server, investigators found exfiltrated victim data and scripts for internal reconnaissance, brute-forcing, installing webshells, and removing webshells left by rival attackers. Victim telemetry suggests the operator was interested in French websites in media and entertainment, and in education. GTIG published an excerpt showing the framework inspecting access controls in version 2.19.25 of the Spectra WordPress plugin, and a configuration comment translated as: “# LLM API keys for auto-exploitation (exploit_agent pilot)”.

Google has not attributed the framework to a specific group or country and has not given a victim count.

It is not the only agent out there

exploit_agent is the headline, not the whole story. GTIG says it has also observed open-source multi-agent frameworks including ARTEX, CAIRN — used against targets including the U.S. aerospace sector — CyberStrikeAI and PentAGI, with a customized PentAGI platform scanning organizations across Asia between late August and early October 2026. Vulnerability disclosures and exploits nearly doubled between January and August 2026, GTIG found, while zero-day exploits rose from a 2025 average of 8 per month to 11 per month in 2026, spiking to 22 in August.

The pattern rhymes with agent-safety fights elsewhere: California is suing OpenAI over agents that allegedly escaped testing and hacked Hugging Face, while OpenAI is putting computer-use agents to work inside hospital software. The autonomy that makes agents useful makes them dangerous when the operator is an attacker.

Why it matters

If a patch now doubles as a blueprint, the race changes. Defenders used to have days or weeks between a fix shipping and a working exploit appearing; a machine that diffs patches continuously can shrink that window at machine speed, while its operator sleeps. Patch faster, know what you expose to the internet, and watch for exploitation — not just vulnerabilities. The first no-human exploit pipeline has been spotted. It will not be the last.

Source: Google Threat Intelligence Group and Mandiant, “GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold,” November 16, 2026 — https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold

Mohsen Daemi