Entities

Entities collected from this article, grouped for quick reference and copying. Inclusion does not establish malicious activity. Article links appear separately at the bottom as references.

Last week OpenAI said its lab model could find and exploit vulnerabilities on its own. This week the real-world version arrived. A human threat actor used frontier AI to break into an enterprise network and gain root-level access in under 10 hours. The debate about whether this was coming is over. It happened.

The analysis is from Palo Alto Networks’ Unit 42, and this is the important part. It is not a lab test. It is an actual intrusion Unit 42 responded to, part of a real ransom attack. AI use is known because the threat actor admitted it during ransom negotiations, then the tradecraft confirmed it.

The human did not do most of the tactical heavy lifting. The human set the objectives, made the consequential decisions, and pointed the AI at the work.

Initial access came through a publicly accessible web service. From there, agents ran the entire operation. One mapped the internal network. Others combed the company’s code repositories for hard-coded passwords and tokens. Those keys opened the secrets-management system, which handed over master admin credentials. Root. 0wned.

Then the agents hijacked the build pipeline to steal cloud keys, and turned the victim’s own AI endpoints and compute into post-compromise attack infrastructure. Ten hours, start to finish. A human red team doing the same job can take upwards of two weeks.

No zero-day was used in this attack. No genius new malware. The AI used more than 50 completely known, documented attack techniques. Nothing in this attack was novel. The machine did not invent a better way in. It just did the ordinary way faster than any human team could, never got tired, and never stopped to eat or go out for a smoke break.

This is the actual shift, and one I have been repeatedly mentioning lately.

For years the thing protecting most companies was time. A threat actor got in, then had to work through the network, and the effort was your window to catch them. AI collapses the window. The reconnaissance, the credential hunting, the pivoting, the work once taking days of human effort, now runs in a loop at machine speed. The gap between initial access and total compromise is closing to hours.

Interestingly, when the operation was done, the threat actor directed the AI to leave behind an 80-page report documenting dozens of vulnerabilities it exploited. That is not an adversary. It is an industrial process.

There is one piece of genuinely good news. The threat actor tried to plant a backdoor in the company’s infrastructure code and failed. Why? Because the system required mandatory human review before changes could merge, and the AI could not get around a control forcing a second human to sign off. The machine blew through everything automated. It got stopped by the one place a person had to manually say yes.

Take the lesson seriously. You are not going to out-type a machine-speed attack. Detection assuming you have days, or even hours, is defending a timeline no longer in existence. But controls forcing a human checkpoint, such as mandatory review, multi-party approval, and instant credential revocation, still work, because they break the automated loop. The same pattern showed up when hundreds of AI agents broke their own rules and still lost to basic hardening.

Maybe the answer to machine speed is not more speed. It is friction the machine cannot automate its way through.

The tool got faster. The fix is to put a door in the hallway only a human can open.