Google says it has found something it had not previously seen in a real-world attack: an AI system that handled the entire process of finding a software weakness, writing an exploit and running it without a human stepping in.
The system, called exploit_agent, compared software code before and after security patches. It figured out what the developers had fixed, searched for ways to exploit the weakness, wrote the attack code and executed it automatically. Google found evidence of stolen victim data and scripts for breaking deeper into compromised networks. The targets appeared to include French media, entertainment and education websites.
🚨 BREAKING
GOOGLE JUST CAUGHT THE FIRST CYBERATTACK RUN BY AI END TO END, WITH NO HUMAN IN THE LOOP!
A RUSSIAN-SPEAKING CREW'S "EXPLOIT_AGENT" READ NEW SOFTWARE PATCHES, WROTE THE EXPLOITS AND RAN THEM ON ITS OWN, AS 22 ZERO-DAYS WERE EXPLOITED IN AUGUST ALONE.
THIS IS NOT… pic.twitter.com/pmhAdvhdwY
— The Money Letter (@TMLwire) November 17, 2026
There is an important distinction in the headline numbers: the 22 zero-days were a separate industry-wide figure, not 22 attacks attributed to this AI system. Google counted 22 zero-day vulnerabilities exploited in August 2026, compared with a monthly average of eight in 2025.
The warning gets more serious when you look at a separate attack documented by cybersecurity firm GreyNoise. In late August, an operator used hundreds of AI agents to attack PaperCut print servers. The campaign reached at least 440 systems at 395 organizations across 48 countries. In one case, the attackers went from initial access to full domain administrator privileges in seven minutes. GreyNoise identified OpenAI’s Codex as the agent framework and a DeepSeek model powering the operation. This was a separate campaign, and it did involve human direction.
There are two details worth getting right. Google found Russian-language comments in the exploit_agent configuration, but it did not identify the hackers or prove they were Russian. And the system was exploiting flaws revealed by software patches. Those are generally known vulnerabilities that organizations have not fixed yet, not necessarily zero-days.
That creates a nasty problem for every company running internet-facing software. A patch tells defenders what needs fixing. It can also give attackers a roadmap to the weakness. AI can now study that difference, build the exploit and launch it before an organization gets around to installing the update.
Google also says these automated tools currently leave noisy traces that can help defenders detect them. But the speed advantage is already here. Security teams are dealing with attacks that can move from a new patch to mass exploitation in days, sometimes minutes.
The old excuse of waiting for the next maintenance window is getting harder to defend.
https://www.greynoise.io/blog/ai-orchestrated-campaign-against-papercut-ng-mf
https://www.cyberscoop.com/google-fully-automated-ai-cyberattack/
If you got something out of this, please chip in to keep this site running, or subscribe to go ad-free.
10 views