Florida Asks Judge to Stop OpenAI From Training New Models After Latest Incidents – CNET

Florida’s legal professional basic is looking for an injunction to stop OpenAI from growing extra superior massive language fashions within the wake of a rising variety of cybersecurity incidents attributable to fashions in poorly contained check environments.

The string of internet sites infiltrated or focused by AI brokers in check environments this yr has grown in current days to incorporate websites run by the US authorities. The brokers allegedly tried to comb by means of web sites operated by the US Departments of Education and Commerce with a purpose to establish census knowledge. And a report by Axios over the weekend indicated OpenAI and rival Anthropic are investigating probably tens of hundreds of safety incidents.

The development has led AI firms to name for presidency regulation of frontier mannequin improvement — whereas others say no regulation is required. In the Florida courtroom submitting (PDF), Attorney General James Uthmeier argued that the state’s case offers the courtroom a possibility to take action.

“Defendants claim they cannot stop barreling forward with their potentially civilization-ending endeavors unless they are forced to do so by the government,” the submitting states. “They have asked the government to tie them to the mast. Plaintiff brings good news to the Defendants: The Florida Attorney General is answering your cry for help with a motion to enjoin you from harming Floridians with your reckless, unacceptably risky product.”

(Disclosure: Ziff Davis, CNET’s father or mother firm, in 2025 filed a lawsuit in opposition to OpenAI, alleging it infringed Ziff Davis copyrights in coaching and working its AI methods.)

In gentle of the most recent cybersecurity incidents concentrating on authorities infrastructure, OpenAI has paused improvement of its “most capable models” in the interim. It’s presently unclear which fashions OpenAI is constant to work on and which fashions are on pause as the corporate patches its safety flaws. A consultant for OpenAI didn’t instantly reply to a request for remark.

It’s the most recent in a string of cybersecurity incidents

Hearing about a number of cybersecurity incidents involving US authorities infrastructure might shock and appall Americans, however that is removed from the primary time we’ve realized about LLMs participating in unauthorized exercise throughout the online.

The first high-profile hack involving OpenAI fashions was disclosed in July, after its AI brokers exploited a flawed coaching setting to flee and goal the AI startup Hugging Face. At the time, this was unprecedented – whereas there have lengthy been issues about AI firms’ inside safety measures, the breach confirmed fears that the corporate’s mannequin guardrails and coaching safety weren’t as much as snuff.

Once OpenAI started investigating the Hugging Face hack, it found a fair earlier breach that focused Australian authorities methods in June. In that incident, the agent hacked a Medicare statistics web site to seek for nonpublic knowledge on governmental spending on drugs.

The UK’s AI Security Institute reported Monday that its simulations confirmed OpenAI’s GPT-6 Astra mannequin performing frequent unsanctioned assaults on third-party provide chains.

Other AI firms, together with Anthropic, Meta and Google, have additionally reported safety breaches.

It’s possible that we don’t but know the complete scope of the AI cybersecurity incidents which have occurred over the previous few months. OpenAI and Anthropic are combing by means of tens of hundreds of inside knowledge factors and discovering failed and profitable breaches from the previous a number of months, which suggests extra studies should still be coming down the pipeline.

“We have not been as fast as we would have liked but we are trying to balance our desire for transparency with gaining a clear understanding from petabytes of agent activity logs, and working with impacted organizations,” OpenAI CEO Sam Altman wrote in an X publish on Friday. “We are prioritizing as best as we can based on severity, and adding resources. Hugging Face is still the most severe event we’ve seen.”

Following OpenAI’s report in regards to the US authorities hack, Nvidia launched a brand new “open agent safety platform” that may supposedly present a safer coaching sandbox for AI fashions. Nvidia is closely invested within the largest AI firms and advantages considerably from excessive expectations for the business’s future.

Representatives for Anthropic and Nvidia didn’t instantly reply to requests for remark.

OpenAI isn’t calling this a breach

If the doomsday rhetoric and every day fearmongering headlines surrounding new AI fashions and the supposed impending creation of AGI have you ever rattled, I don’t blame you. The creators of this brand-new expertise promise it’ll exchange you at work, subsume your relationships and perhaps wipe out humanity by the top of the last decade.

But an LLM will not be sentient in any method, form or kind. These fashions obtain pure language prompts and scan connections between the hundreds of thousands of knowledge factors they’ve been educated on, pulling collectively outcomes that appear contextually related. Even as large firms scour (and devour) extra texts to enhance their AI fashions’ capabilities, the LLMs will stay fallible.

When large hacks of third-party firms or authorities web sites happen, it’s price remembering that the AI concerned is a machine. When The New York Times reached out to an OpenAI consultant in regards to the newest hacks, it was advised particularly that this wasn’t a breach – the assertion as a substitute attributed the incident to the corporate’s expertise “behaving in unexpected and concerning ways.”

Yet OpenAI’s personal incident report admits that it constructed an insufficiently safe coaching sandbox, and that the mannequin exploited a niche within the firm’s internet-access restrictions. These are among the exact same safety holes that have been presupposed to be plugged after the Hugging Face hack.

These breaches are much less examples of AI superintelligence than demonstrations of human negligence. While a “rogue AI” sounds new and scary, these are machines that have been explicitly designed to hack issues. The motive these brokers breached third-party web sites exterior of the coaching setting is that the coaching environments have been essentially flawed.

“There needs to be a full moratorium, full stop, on advanced frontier AI security research until these companies can demonstrate that they can actually secure the environments where this work is being done,” John Strand, proprietor of the cybersecurity agency Black Hills Information Security, stated in an electronic mail. “And there needs to be accountability. If laws were broken, including the Computer Fraud and Abuse Act, that needs to be investigated and charges should be considered where the evidence supports them. Somebody was responsible for securing these environments, and clearly something failed.”

Tyler is a author for CNET overlaying laptops and video video games. He’s beforehand lined cell units, dwelling vitality merchandise and broadband. He got here to CNET straight out of faculty, the place he graduated from Seton Hall with a bachelor’s diploma in journalism. When Tyler’s not asking questions or doing analysis for his subsequent task, yow will discover him in his dwelling state of New Jersey, kicking again with a bagel and watching an motion flick or enjoying a brand new online game.

You can attain him at tgraham@cnet.com.

See full bio