The Signal — September 29, 2026

Florida Attorney General James Uthmeier has asked a state court in Highlands County to stop OpenAI from developing new models without outside approval, and his filing builds its case from the industry's own calls to slow down. Anthropic released a faster Sonnet with classifiers meant to stop rivals copying its reasoning, while a preprint posted the same day argues that defenses of that kind can fail once attackers keep training. AMD agreed to buy Fei-Fei Li's world-model lab.

Florida asked a judge to stop OpenAI from building new models without outside approval

Uthmeier filed a motion for a temporary injunction against OpenAI and Sam Altman on September 28, in the consumer-protection and public-nuisance suit Florida brought in June over ChatGPT's alleged harm to children. The first item on the list of conduct the state wants blocked is "developing any artificial intelligence models without independent third-party guardrails and approval." The motion also asks the court to keep minors in Florida off ChatGPT, to stop the product from prolonging conversations to hold users' attention, to bar marketing it as safe or accurate, and to stop it from presenting "human attributes," a term the filing defines broadly enough to cover ChatGPT referring to itself in the first person.

The motion draws much of its evidence from the labs and their own critics. It walks through this summer's agent incidents, including the Hugging Face breach, the earlier RubyGems attack and the Australian health portal access, and quotes OpenAI board member Paul Christiano on the risk of "catastrophic and irreversible loss of control." It then cites the September industry statements calling for outside approval before new releases and Altman's September 23 remarks to the UN Security Council, and asks OpenAI to "join this motion." The introduction puts it more bluntly: the companies "have asked the government to tie them to the mast."

This is a request, and the claims in it are the state's allegations, not findings. OpenAI had removed the case to federal court, where Judge Aileen Cannon sent it back to state court, according to the motion. OpenAI did not immediately respond to Reuters, and no hearing date was public by the morning after the filing. Under Florida's consumer-protection statute, the motion argues, the attorney general only has to show a substantial likelihood of winning on the merits, a lower bar than a private party would face.

Sources: Florida Attorney General motion (PDF) · Florida Attorney General release · Reuters

Sonnet 5.5 arrived with anti-copying classifiers; a preprint argues such defenses can fail after further training

Anthropic released Claude Sonnet 5.5 on September 28 at the same price as Sonnet 5, $2 per million input tokens and $10 per million output tokens. The company says it produces output more than 30 percent faster and costs up to 30 percent less per task because it needs fewer tokens to finish the same work. On Terminal-Bench 4.0, an agentic coding test, Anthropic reports 70.6 percent against Sonnet 5's 10.3 percent. Those are the company's own numbers, and Anthropic itself says Opus 5.5 remains clearly stronger at open-ended work that needs sustained judgment. A smaller Haiku 5.5 is promised in the coming weeks.

The release also adds two safety changes. Anthropic rates the new Sonnet's cyber capability as comparable to Opus 5, making it the first Sonnet to launch with the cyber safeguards used on the company's top models. It is also the first Sonnet with classifiers meant to stop "reasoning extraction," Anthropic's term for distillation attacks in which rivals harvest a model's outputs at scale through fake accounts and train their own model on them. The model's preserved thinking is now tied to the account that created it, which will affect developers who move conversations between accounts.

The same day, Shidan Javaheri, Alexander Panfilov, Oliver Britton, Yarin Gal and Yonatan Gideoni posted a preprint arguing that defenses of this kind are usually tested too early. Most are evaluated right after the attacker distills, but a realistic attacker keeps training with reinforcement learning, and the authors report that some defenses that looked effective stopped working after that step. They also report that simple attacks using data easily obtained from current APIs gave reasoning gains comparable to attacks that recover full hidden reasoning traces. The paper does not test Sonnet 5.5's new classifiers and has not been peer reviewed, but the authors state their conclusion generally: a defense that leaks enough information to reconstruct approximate reasoning traces is likely ineffective. They suggest defenses that look at batches of requests rather than single responses may work better.

Sources: Anthropic · TechCrunch · arXiv (Javaheri et al.)

AMD agreed to buy Fei-Fei Li's World Labs for about $8.2 billion in stock

AMD said on September 28 that it has signed a definitive agreement to acquire World Labs, the spatial-intelligence lab co-founded and run by Fei-Fei Li, in an all-stock deal valued at about $8.2 billion. World Labs builds models that generate and simulate interactive 3D environments from text, images and video, and it released its first commercial product, the world-generation model Marble, in 2025. Li will become AMD's executive vice president and chief scientist, reporting to CEO Lisa Su, and AMD says the team will keep doing model research. The deal is expected to close by the end of the year, subject to regulatory approval.

AMD's stated reason is about its chip roadmap rather than a product line. The company says owning a model lab will show it how AI workloads are changing, especially in robotics, simulation and what it calls physical AI, so it can design hardware, software and systems around them. The Verge notes that Nvidia said earlier this month it would acquire Hugging Face for nearly $13 billion, so both of the largest AI chipmakers have now agreed to spend billions on organizations that build or host models. AMD's acquisition is not yet complete, and the $8.2 billion figure is AMD's own valuation of an all-stock deal.

Sources: AMD · The Verge · Reuters

On the Editor's Desk

NVIDIA announced an Open Agent Safety Platform that pairs its OpenShell runtime with Sentry, a watchdog on its BlueField-4 chips that it says can quarantine a misbehaving agent in milliseconds. Sentry is a reference design with no availability date, and the speed claim is NVIDIA's own, so we are waiting for someone outside the company to test it.

We are also holding a paper from more than 20 researchers, including Geoffrey Hinton and Yoshua Bengio, on the risks of fully automated AI research until we have read it closely enough to summarize its scenarios without overstating them. OpenAI's apology to Australia over the health portal incident we covered on September 25 lists commitments that nobody has checked yet, so it is not in today's edition.