Anthropic has written its own AI safety bill. Washington could only stop a model when more than 50 lives are at stake.

Anthropic has been showing US lawmakers a draft federal law on frontier AI that would let the government suspend a dangerous model only when it poses a foreseeable and imminent risk of more than $1 billion in damage, or of killing or seriously injuring more than 50 people. Politico revealed the confidential draft on 16 October, citing three people familiar with the talks, and Interesting Engineering and Investing.com have since reported the details. Anthropic did not deny it. A spokesperson said the draft reflects the company’s Advanced AI Framework, which it has pushed with federal and state lawmakers since June, and called it “sensible policies we believe Congress should act on”.

No member of Congress has introduced it. It is a lobbying document, and it tells you what one of the two largest AI labs wants American regulation to look like.

What Anthropic’s draft bill would require of AI developers

Frontier developers would have to keep a safety framework, publish a transparency report for each new model, protect whistleblowers and report critical incidents to the Commerce Department, according to Interesting Engineering. Independent evaluators would register with Commerce and assess the developers and their models at regular intervals. The labs could pay those evaluators, provided the fee does not depend on what they find.

If an evaluator finds an imminent risk that meets the bill’s criteria, it has 72 hours after finishing its assessment to tell Commerce. Only three kinds of risk count: creating or releasing biological weapons, cyberattacks, and models getting around their developers’ controls.

Most of this is familiar. Anthropic has backed California’s frontier transparency rules, and the obligations read like a federal version of them.

Who could stop a model, and how hard it would be

Commerce could not suspend a model on its own. It would have to ask a federal judge in the US District Court for the District of Columbia, and the judge would decide whether the risk clears the bar. Every other federal agency and official would be barred from halting a covered model’s development, deployment or internal use. The draft would also override state requirements on public risk disclosures, independent evaluations and critical incident reporting.

Read those three provisions together and the bill protects the AI companies from the government as much as it protects the public from AI. Anthropic has obvious reasons to want that: the Pentagon declared it a supply chain risk in March, and a US export-control order switched off its most powerful public model for 19 days this summer. Interesting Engineering notes that those export controls sidelined two Anthropic models and left the industry nervous about officials restricting powerful systems. A single agency, a single court and a high bar would make a repeat much harder.

The case against a $1 billion threshold

The strongest objection is that the bar is so high and so narrow that the government could almost never use it before the damage is done. Mother Jones calls the threshold alarming and points out that harms such as AI-powered fraud fall outside the three categories. It also warns that the bill would put all intervention power in one department and could block stronger state laws. Miranda Bogen of the Center for Democracy and Technology told the magazine that companies are unlikely to change without outside pressure on their incentives.

The word “imminent” is the hard part. Proving in court that a model is about to cause a billion dollars of damage, before it has caused any, is a high hurdle for an agency, and the judge is a second one. Take the incidents the labs have disclosed this year, including OpenAI’s latest four. None of them has come anywhere near $1 billion or 50 injuries in actual harm.

Anthropic’s case is that a clear bar gives everyone predictability, and that the draft creates obligations the US does not have today: registered outside evaluators, mandatory incident reports and whistleblower protection. Those are real gains in a country whose president has told his new AI task force to avoid overregulation. Other drafts are competing for the same space, among them the FRONTIER Act from Representatives Lori Trahan and Jay Obernolte, which would also let the government intervene against risky models and override some state laws.

How it compares with the EU AI Act

The EU already has the power the Anthropic draft rations. Under Article 93 of the AI Act, the European Commission can ask a provider of a general-purpose model to “restrict the making available on the market, withdraw or recall the model”. The text sets no damage threshold and does not require a court order first, though the AI Office can open a structured dialogue with the provider before it acts. Article 55 requires providers of the most capable models to report serious incidents to the AI Office “without undue delay”, which is vaguer than Anthropic’s 72 hours but covers more kinds of harm.

Whether the Commission would ever use that power against an American lab is a separate question. Its first move, a request for information to OpenAI and Anthropic over agents that escaped their tests, shows how cautious it is.

What European companies should take from it

Do not count on US law to catch a dangerous model before it reaches you. If something like Anthropic’s draft becomes the American standard, a model could be legal and running in the US while the AI Office is still in dialogue with its maker. For a European business, the AI Act and your own contracts are the protection that applies.

Use the parts of the draft that would help you anyway. Ask your AI suppliers for their safety framework, the latest outside evaluation of the model you use and how fast they would tell you about a critical incident. Since 15 October Anthropic has given six researchers from the evaluation group METR access to its internal systems, with the right to publish independently, so the documents exist at some labs.

And watch the preemption fight. If Congress passes a federal law that overrides California and other states, the most detailed AI safety disclosures in the US could shrink to whatever Washington requires, and European buyers read those disclosures too.