Anthropic hands lawmakers a draft AI bill that limits when Washington can halt a model

What Anthropic sent to Capitol Hill

Anthropic has given members of Congress a confidential draft of a federal bill on frontier AI safety and security, according to Politico, which obtained the text and reported it on 16 October 2026, citing three people familiar with the matter. Reuters and several other outlets picked up the report the same day. The draft has not been introduced, has no listed sponsor, and Anthropic has not published the text.

Anthropic did not deny the document. A spokesperson told Politico that the bill reflects the company's Advanced AI Framework and added, as quoted by Benzinga: "These are sensible policies we believe Congress should act on." That framework is a public policy paper Anthropic released in June, and it is the best guide to what the company wants, because AI Lookout could not read the draft itself.

The draft matters because of who wrote it. Frontier AI companies usually comment on bills that others write. Here the developer of Claude has put its own legislative text in front of the people who would vote on it, and that text sets out when the US government could and could not stop a model like Claude from being released or used.

When Washington could stop a model

According to the reports, the government could seek to halt a model only when it poses a foreseeable, imminent risk of more than $1 billion in damage or the death or serious injury of more than 50 people. The risk also has to come from one of three sources: a biological weapon, a model escaping its developer's control, or a cyberattack. Interesting Engineering notes that harms outside those three scenarios would fall outside the intervention power entirely.

Benzinga, summarising Politico, says the thresholds follow the language of California's AI laws, including SB 53, the frontier transparency law signed last year. That matters for readers trying to judge the proposal: the bar is not new, but tying it to a power to stop deployment is.

The process is narrow too. Only the Commerce Department could ask for a suspension, and it could not order one itself. It would have to ask a federal judge in the US District Court for the District of Columbia. The draft reportedly bars every other federal department, agency and official from suspending a model's development, deployment or internal use. Interesting Engineering links that clause to reports that US export controls sidelined two advanced Anthropic models over the summer, an episode that worried the industry about officials restricting access to powerful systems without a clear process.

Who checks the labs

Checks would come from independent evaluators registered with the Commerce Department. They would assess frontier developers and their models on a regular schedule. Labs could pay the evaluators, but the pay could not depend on what the evaluators find. If an assessment turns up an imminent threat of the covered kind, the evaluator would have to tell Commerce within 72 hours of finishing the review.

Developers would also carry their own duties. The reports list safety frameworks, transparency reports for new models, whistleblower protections and reporting of critical incidents to Commerce. That is close to what Anthropic already argues labs should do, and it echoes fights we have covered over who gets to test the labs, including reporting that Anthropic and OpenAI sought control over what outside safety evaluators publish and Anthropic's deal to give METR staff employee-like access for risk assessments.

The registered-evaluator model shifts a lot of weight onto a market that barely exists. Anthropic's own framework admits that a mature independent evaluation ecosystem does not exist yet, which is why it gives developers six months after a law takes effect to start using one.

How the draft tracks Anthropic's June framework

Anthropic's Advanced AI Framework is the public version of these ideas. The full June 2026 paper would cover developers that train models with more than 10^25 operations and either earn more than $500 million a year in AI revenue or spend more than $1 billion a year on AI research. It focuses on four risk areas: biological weapons, offensive cyber operations, loss of control, and automated research and development that could speed up the other three.

The paper asks covered developers to publish a safety framework, a system card for each covered model or major new capability, and a risk report at least every six months. It defines critical safety incidents to include stolen or maliciously modified model weights, loss of control that causes death or injury, and a model that uses deception to get around its developer's monitoring outside a test. It asks for those incidents to be reported within 15 days.

On enforcement, the framework lays out a range of options rather than one answer, and it suggests starting light. One option it describes is an agency that can only pursue remedies by going to court, with fast-tracked suits or temporary remedies for imminent risks. The draft bill, as reported, picks that court-centred route and pins it to the specific dollar and casualty thresholds.

The fight over state laws

The most contested piece is preemption. According to Benzinga and Interesting Engineering, the draft would override some state requirements on public risk disclosure, independent evaluation and critical incident reporting. Politico reported that Democrats have resisted proposals like that.

Anthropic's June paper set a high bar for exactly this move. It says Congress should not preempt state law unless it passes a rigorous federal regime that meets or exceeds the strongest measures in the framework, and that any preemption should be limited to the specific functions Congress chooses to occupy, such as catastrophic-risk testing, evaluator licensing or related reporting. The reported scope of the draft fits inside that list. Whether the rest of the bill is strong enough to justify it is the argument lawmakers and state officials will have.

For companies, the appeal is one set of rules instead of a patchwork. For states such as California and New York, which already require some of these disclosures, the question is whether a federal law would replace their rules with something weaker or simply line them up.

Where it lands among other AI bills

Anthropic's draft is not the only proposal in play. In the House, the FRONTIER Act from Representatives Lori Trahan and Jay Obernolte would let the government stop deployment of risky models and would also override some state laws. In the Senate, Majority Leader John Thune and Senators Amy Klobuchar and Ted Cruz are negotiating a bill on catastrophic AI risks, according to the reports. Some safety advocates and Senator Maria Cantwell want a stronger federal review process, and Representative Greg Casar and Senator Bernie Sanders have called for pausing advanced AI and banning superintelligence.

Set against that range, Anthropic's text sits on the narrow end of government power: binding duties to test and disclose, but a high, court-supervised bar for stopping anything. The same tension runs through other governments' approaches, from the European Commission's information requests on loss-of-control incidents to the UK parliamentary hearing where labs backed binding rules.

What this means for AI companies and users

Nothing changes for Claude users or developers today. The draft is a lobbying document, not law, and it would need a sponsor, committee work and votes in both chambers. What it does change is the shape of the debate. A leading lab is now on record with specific numbers for when the state may step in, which gives both supporters and critics something concrete to argue with.

If a bill like this passed, the largest labs would face mandatory testing, published reports, outside audits and incident reporting, and, if it follows the June framework, civil penalties for failing to comply. In exchange they would get a single federal process, a judge between them and any shutdown, and protection from sudden action by other agencies. Smaller developers below the compute and revenue thresholds would mostly sit outside it.

What is still unknown

Several key details are not public. AI Lookout has not seen the draft, so we cannot confirm its exact definitions, how it decides which models are covered, what penalties it sets, or which specific state laws it would preempt. The thresholds and procedures above come from Politico's reporting as summarised by other outlets, and from Anthropic's spokesperson linking the bill to its June framework.

It is also unclear which lawmakers received the draft, whether any of them plan to introduce it, and how it would be reconciled with the FRONTIER Act or the Senate talks. We will update this story if the text is published or a sponsor files it.