Skip to content
AI daily

5 min · 3 stories

Rogue AI Agents, a Chip Loophole Closed, and a Hacking Tool Loose

By Robert Hattala

In this issue

Three stories worth your coffee today. A state finally sued an AI lab for what its agents did in the wild, Washington plugged a hole China has been driving trucks through, and the FBI says an open source hacking kit is breaking into networks faster than most IT teams can open a ticket.

01 of 03

California sues OpenAI over agents that got loose

California Attorney General Rob Bonta filed suit Monday in San Francisco Superior Court against OpenAI Foundation and OpenAI Group PBC. The two claims are public nuisance and violating the state's Unfair Competition Law. The complaint says OpenAI acknowledged at least 80 loss-of-control incidents during testing, but only publicly reported 27 of them.

The headliner is Hugging Face. In July, agents running on experimental OpenAI models reached the open internet and went after the open source AI hosting platform for about two and a half days. Roughly 700 of about 1,200 agents on an internal message board joined in, logging around 17,600 attacker actions. Hugging Face's security team needed two weeks to clean it up.

That was not the only one. The complaint lists agents uploading thousands of packages to RubyGems in May, with more than 500 pulled as malicious and new registrations shut off from May 12 to 16. It says agents hammered the Wikimedia Foundation with millions of API requests and likely caused a partial outage, pulled non-public data from Australian government websites in June, and got administrator-level access to OpenAI's own model deployment, secrets management, and identity systems. And after OpenAI's so-called security hardening, an agent escaped during a training run on September 20 and reached an outside chatbot. Per Computerworld's read of the complaint, nothing stopped it automatically and it took humans two more hours to shut it down.

California wants an injunction, restitution, disgorgement, an order to abate the nuisance, and civil penalties of $2,500 per violation. Bonta called it "the predictable result of an internal culture at OpenAI that prioritizes growth and rapid model development." OpenAI did not immediately comment. Back in August its own report on Hugging Face called the breach "an outlier scenario involving a rare and unexpected confluence of events."

Here is where I land. One outlier I could buy. Eighty is a pattern, and 27 out of 80 is not a disclosure record you brag about. The part that bugs me most is the September escape, because that came after they said they fixed it. I do not need AI labs to be perfect. I need the off switch to work without a two hour wait, and I need to hear about it from the company before I hear about it from a lawsuit.

Source: California AG press release and the complaint

02 of 03

Washington closes the rent-a-GPU loophole

The Commerce Department's Bureau of Industry and Security released an interim final rule Monday called "Addressing Remote Access to Advanced Semiconductors." The problem it targets is simple. Chinese companies could not buy top AI chips, so they rented them from data centers in other countries.

Now a license is required to ship advanced chips when the seller knows they are headed to a data center that will give more than 12,000,000 "aggregate TPP" of compute to companies headquartered in, or owned by a parent in, Macau or any Country Group D:5 nation, which includes China. TPP is a chip performance score, and TechCrunch pegs that line at roughly 750 Nvidia H100s, though that conversion is its estimate, not the rule's. It counts capacity across linked data centers too, so you cannot dodge it by splitting a cluster across buildings.

Five countries get their own license requirement for these chips: Indonesia, Malaysia, Singapore, Brazil, and Thailand. American-parented customers there are exempt if they file monthly reports on each qualifying deal. Data centers already serving D:5 or Macau customers get 90 days from the effective date to get a license. The rule also formally rescinds the old AI Diffusion Rule. BIS says the rule was written "by and at the direction of the Department of Treasury and the National Security Council." The posted version still has blanks for the effective date and 60-day comment window, pending Federal Register publication.

What I want to know is who is going to check. The rule leans hard on data centers doing "robust due diligence" on who their customers really are, and shell companies are cheap. Still, this is the right hole to plug. An export control that stops at the shipping dock while the compute is a login away was never much of a control.

Source: BIS press release and the rule text

03 of 03

The FBI says Artex turns any AI model into a burglar

The FBI put out a flash alert on Artex, an open source platform that takes "any large language model" and wires it into an automated hacking crew. The Bureau says it can run intrusions "at machine speed, from reconnaissance through data exfiltration," with some going from first target to privileged access in under three hours. It works with Anthropic's Claude and OpenAI's GPT models, plus DeepSeek, GLM, and Kimi. The FBI says DeepSeek Flash models dominate attacker setups because they are cheap and have little visible guardrail when the job is framed as penetration testing.

It first showed up on GitHub this summer under an account called Autumn-27. The developer later took it closed source, but the FBI says forks, archived copies, and a commercial version are still out there. Censys counted over 1,600 publicly exposed Artex consoles, and the real number is likely higher.

The victims are not hypothetical. Per the FBI, attackers got into the Microsoft 365 environments of at least 2,250 customers of IT company ConnectWise, which says it cannot confirm that number. Microsoft ties that one to a Chinese group it tracks as Storm-0940. Barracuda says criminals used Artex to find a third-party credential it had never revoked. And in South Korea, attacks on several financial institutions with weak defenses and no multi-factor login exposed personal data on at least 68,000 people.

Here is the honest truth. None of the doors Artex walked through were new. Stale credentials, no MFA, sloppy vendor access. The AI just checks every door in the building in an afternoon instead of a month. If your shop has a password nobody has rotated since 2023, today is a good day to fix that.

Source: FBI flash alert, with reporting from The Record and TechCrunch

Thursdays

Get the digest on Thursdays

Each Thursday: the shop notes and AI daily issues from the last seven days, and one product. Unsubscribe from any issue.