Anthropic, the San Francisco company behind the Claude chatbot, has drafted a federal AI safety bill that would limit when the government can stop an advanced model, a report on the unpublished text says. Action would need an imminent threat of mass harm and a federal judge’s approval. Anthropic has not released the text, and the bill has not been introduced in Congress.
Details of the confidential draft come from a Washington news report based on a copy of the text and three unnamed people. Those sources say Anthropic has been showing the draft to lawmakers, and no sponsor has been named. The company says it reflects its Advanced AI Framework, a policy paper it has promoted with federal and state lawmakers since June.
“These are sensible policies we believe Congress should act on,” a spokesperson said.
Alongside the limits on halting models, the largest AI developers would take on new duties under the draft. They would publish safety frameworks and transparency reports, accept regular checks by independent evaluators, protect whistleblowers and report serious safety incidents to the Commerce Department. The bill would also override some state rules in the same areas.
A High Bar For Halting A Model
To trigger action, a model would have to present a foreseeable and imminent risk of more than $1B in damage, or of killing or seriously injuring more than 50 people. That risk would also have to come from one of three sources: creating or releasing a biological weapon, a model escaping its developer’s control, or a cyberattack.
Even then, the Commerce Department could not act alone, according to the reported text. It would have to ask a federal judge in the District of Columbia to order a suspension. No other part of the federal government, from departments down to individual officials, could pause a model’s development, release or use inside the company that built it.
Independent Evaluators And Reports
Much of the checking would fall to outside evaluators, independent testers registered with the Commerce Department. They would assess frontier developers and their models on a regular schedule. The companies could pay them, but not in a way that depends on what the evaluators find.
If an evaluator spotted an imminent risk of the kind the bill covers, it would have to tell the Commerce Department within 72 hours of finishing its review. Developers would separately report critical safety incidents to the department, alongside publishing their own safety frameworks and transparency reports for new models.
Roots In Anthropic’s June Framework
Anthropic first set out many of these ideas in its public Advanced AI Framework. Dated June 2026, the paper is written mainly with the federal government in mind. It would cover companies that train a model using more than 10²⁵ floating-point operations (FLOP), a measure of raw computing, and also earn over $500M a year from AI or spend over $1B a year on AI research and development.
Those developers would publish risk reports at least every six months, release system cards, public write-ups of a model’s abilities and testing, for major new models, and report critical safety incidents within 15 days. The paper suggests civil penalties that grow with a company’s global annual revenue.
To guard against government overreach, the framework lists options that can be combined. One would make the agency go to court for remedies such as restricting a model already in use, rather than impose them itself. Another would let developers challenge remedies through fast-track judicial review.
The framework names four risk areas where the reported draft names three, adding automated research and development that could speed up the other risks. A second part covers wider public defences, such as screening orders for synthetic genes and making critical infrastructure operators patch software on a fixed schedule.
Anthropic had its own clash with federal power earlier this year, when it took two of its latest models offline to comply with a Trump administration export-control directive aimed at keeping them from foreign nationals. The company said the government should be able to block unsafe deployments through a legal process that is “transparent, fair, clear, and grounded in technical facts”, adding: “This action does not adhere to those principles.”
The Fight Over State AI Laws
California’s Senate Bill 53 (SB 53), signed in September 2025, already requires large frontier developers to publish safety frameworks and report critical incidents to the state. Anthropic endorsed it at the time, saying frontier AI safety “is best addressed at the federal level instead of a patchwork of state regulations”, but that AI progress would not wait for Washington to agree.
New York’s Responsible AI Safety and Education (RAISE) Act, signed in December 2025, sets similar duties, and state leaders agreed to amendments bringing it closer to California’s law. California’s law and New York’s amended version both treat a risk as catastrophic if it could kill or seriously injure more than 50 people or cause over $1B in damage, the same lines used in Anthropic’s draft.
Anthropic’s draft would replace state rules on public risk disclosure, independent evaluation and incident reporting with its federal versions. This is pre-emption, where federal law overrides state law. The June framework was more cautious, saying Congress should do this only after passing federal rules at least as strong as the framework’s toughest measures. Without such a law, it says, states should keep full power to legislate.
The White House has pushed for a single national approach. President Donald Trump signed an executive order in December 2025 directing the Justice Department to challenge state AI laws that clash with national policy. In March 2026, the administration sent Congress recommendations calling for national standards that override state AI laws. Anthropic also signed the voluntary White House accord on superintelligence.
Rival Plans On Capitol Hill
Lawmakers already have several proposals of their own. Representatives Jay Obernolte, a California Republican, and Lori Trahan, a Massachusetts Democrat, introduced the FRONTIER Act in July. It would require independent audits and incident reporting from the largest developers under one national standard, and let the commerce secretary suspend or restrict a model posing an imminent catastrophic risk.
In the Senate, Majority Leader John Thune has worked with Senators Ted Cruz and Amy Klobuchar on a bill aimed at catastrophic AI risks. Those talks were reported in July to have stalled over disagreements with Anthropic and with Senator Maria Cantwell. Anthropic pushed for public disclosure of risks and objected to giving the Commerce Department a standing power to seek court orders blocking models.
Cantwell and some safety advocates want a stronger federal review process than these bills offer. Senator Bernie Sanders and Representative Greg Casar have gone further, proposing a ban on developing superintelligent AI.
A Draft Without A Sponsor
For now, Anthropic’s bill is a private draft with no named sponsor, circulating among lawmakers who already have competing plans and remain divided over how much power regulators and states should hold. It would add federal duties for the biggest AI developers while narrowing who in Washington can stop a model, and when. The company’s wider thinking is public in its June framework; the bill text is not.
