For years, security researchers warned that the day would come when an artificial intelligence could find a software flaw, write the attack code, and run it — all without a human being involved at any step. According to Google, that day has now arrived. In its fourth-quarter 2026 AI threat tracker, published on November 16, 2026, the company's Threat Intelligence Group (GTIG) says it has documented the first known case of attackers fully automating vulnerability discovery and exploitation with an AI agent, with no human in the loop.
The case is not a laboratory demonstration. The AI system, which Google calls "exploit_agent", was found running on an exposed staging server in the wild, and the data it had already taken from victim organisations was sitting on the same machine. This is the moment the theoretical risk became a working weapon — and the details matter, because they show exactly how the automation works.
How exploit_agent turns a patch into a weapon
The framework's trick is elegant and, from a defender's point of view, deeply unsettling. Every time a software vendor releases a security patch, it publishes, in effect, a map of the flaw it is fixing: the code before the patch and the code after it. According to reports from The Register, which first covered Google's report, exploit_agent was configured to continuously compare the pre- and post-patch code of popular web application software, identify the vulnerability the patch had closed, develop a working exploit for it, and then execute that exploit automatically — without waiting for a human instruction at any stage.
Google describes the system as a multi-agent framework, wired into multiple frontier and locally hosted large language models, and built with Russian-language tooling. Its configuration files were written in Russian, and Google describes the operator as a suspected Russian-speaking actor. It is important to be precise here: Google has not established that the operation was coordinated by authorities in Moscow, and no government has been blamed for it. What is established is the language of the tooling and the autonomy of the attack.
The staging server where threat hunters found the framework told its own story. Alongside the framework sat data stolen from victim organisations, and a set of post-compromise scripts for internal reconnaissance, brute-forcing passwords, installing webshells — and even removing other attackers' webshells, so the operator's access stayed exclusive. Victim telemetry suggests the operator was particularly interested in French websites in the media and entertainment industries, and in the education sector.
Why Google calls this a threshold, not just another incident
Google's report is titled "Agentic Threat Activity Crosses a Significant Threshold", and the title is deliberate. Just two months ago, the same group said it had not seen fully autonomous attack pipelines deployed in the wild. The change between then and now is the difference between an AI that helps a hacker and an AI that is the hacker.
On Google's own maturity scale, exploit_agent sits at Level 5 — meaning the AI handles most stages of the attack, with a human at most setting the system up. The practical consequence is speed and scale: an automated system can watch every patch from every major vendor, test every one for a usable flaw, and strike in the window between a patch being published and a victim installing it. That window — often days, sometimes hours — is exactly where most organisations are slowest.
The report also documents a wider shift. Google says attackers have moved beyond manually prompting a chatbot during intrusions and are now deploying autonomous, multi-agent offensive frameworks to carry out multi-stage attacks. It names other frameworks in the same ecosystem — ARTEX, CAIRN, CyberStrikeAI, HexStrike AI and PentAGI — and the FBI has separately warned that the ARTEX penetration-testing framework can take some intrusions from first contact to privileged access in under three hours.
What Google expects next — and what defenders should do
Google's researchers estimate that within about three months, threat actors will be able to launch unsophisticated attacks that need little or no human interaction and rely mostly on multiple AI agents. Their projection is that sophisticated, fully autonomous agentic attacks are likely to emerge within the next six months. That is not a distant forecast; it is a deadline measured in weeks.
For anyone running a website or a business network, the advice has not changed in kind, but it has changed in urgency. Patch quickly, because the gap between a patch's release and an automated exploit is now measured by machines, not by humans reading advisories. Keep internet-facing systems off the public internet when they do not need to be there — exploit_agent was found on an exposed staging server, the kind of forgotten machine that exists in thousands of organisations. And treat patch notes as sensitive: the same diff that tells defenders what to fix tells an AI exactly where to strike.
The broader lesson is the one this blog has tracked before — from the North Korean FaceStealer deepfake campaign to the Windows SignDrop zero-day. AI is no longer only a tool that attackers use; it is becoming the attacker itself. The first fully automated cyberattack was found because its operator left a server exposed. The next one may not be so careless — and the defence that matters most is the unglamorous one: patch fast, expose less, and assume the machine on the other side never sleeps.
