Anthropic wants to be regulated. Just not by anyone’s rules but its own.

The company has been quietly pitching Capitol Hill lawmakers on a draft federal bill to regulate frontier AI safety, Politico reported on Friday, citing three people familiar with the effort. The draft would let the government block a model’s release or use only if it poses a foreseeable, imminent risk of more than $1 billion in damages — or of killing or seriously injuring more than 50 people.

Read that threshold twice. It’s not a safety net; it’s a tripwire set so high that almost no model failure in history would have crossed it.

What’s actually in the draft

The bill narrows the grounds for intervention to three categories of catastrophe: the creation or release of a biological weapon, a model evading its developer’s controls, or a cyberattack. Anything short of that — biased hiring tools, deepfake floods, an agent gone rogue on a government website — doesn’t qualify.

Enforcement would run through a new cast of characters: independent evaluators registered with the Commerce Department, tasked with periodically evaluating the labs and their frontier models. Labs could pay these evaluators directly, provided the payment isn’t tied to the outcome of the evaluation. If the evaluators determine there is “imminent enumerated risk” in a company’s frontier models, they must report it to Commerce within 72 hours of completing their assessment.

Then Commerce can ask a judge to step in. A federal judge for the District Court for the District of Columbia could order a model’s suspension. And here is the part Congress will fight over: the draft explicitly prohibits any other US government department, agency, or officer from suspending development, deployment, or internal use.

One pipeline. One judge. One department. Everyone else: hands off.

Anthropic and the Commerce Department did not respond to Reuters’ requests for comment.

Why Anthropic wrote its own AI safety bill

This is a familiar Washington playbook dressed in new clothes. Regulated industries have drafted their own legislation for a century — it’s how banking, telecom, and pharma got the version of oversight they could live with. Anthropic’s version is notably elegant: it gives critics a headline about “blocking dangerous AI” while quietly preempting the messy multi-agency enforcement the industry actually fears.

The FTC is already circling. The commission has reportedly sent civil subpoenas to both Anthropic and OpenAI this week seeking documents on AI safety incidents, privacy policies, and protective measures. A draft that channels all model-suspension power through Commerce and a single federal court would effectively take the FTC’s shutdown option off the table before the fight even starts.

Congress has been circling this all year

Anthropic’s pitch lands in a crowded legislative field. In July, a bipartisan House group introduced the FRONTIER Act, which would set transparency requirements and independent audits for the largest AI developers. Rep. Lori Trahan later released draft language that would hold developers responsible for harm even when they used reasonable care. Senators Amy Klobuchar, John Thune, and Ted Cruz have been advancing a separate proposal that converts voluntary safety pledges into an enforceable duty of care.

None of those bills has made it out of committee. That’s the opening Anthropic is walking through: if Congress is going to write AI rules eventually, the company would rather hand lawmakers the first draft than react to someone else’s.

Why this matters

The numbers in the draft are the story. One billion dollars. Fifty people. Seventy-two hours. Those figures draw a bright line between “dangerous AI” and everything else the industry gets criticized for, and they put that line somewhere most incidents never reach. If this language became the federal standard, it wouldn’t just regulate frontier AI — it would define, in law, which harms don’t count.

That doesn’t make the draft sinister. Anthropic has genuinely invested in third-party evaluation — it recently gave researchers from the evaluation nonprofit METR employee-like access and let them publish findings without a veto. A formal evaluator regime is a real concession. But the company’s version comes with a friendly price: the evaluators get paid by the labs, and the only government body with a kill switch is Commerce, petitioning one judge in D.C.

The rest is politics. Capitol Hill rarely passes a company’s first draft as written. It amends it, strips the self-protective clauses, and adds teeth the industry didn’t want. Anthropic knows that. The pitch isn’t the bill — it’s the opening bid.

Frequently asked questions

What is in Anthropic’s draft AI safety bill?

The draft, pitched to Capitol Hill and reported by Politico on October 16, lets the government block a model’s release or use only for foreseeable, imminent risk of more than $1 billion in damages or 50 deaths or serious injuries — limited to bioweapons, loss of developer control, or cyberattacks. Evaluators would register with the Commerce Department and report imminent risks within 72 hours.

Can the government shut down an AI model under it?

Only a federal judge in the District of Columbia could order a suspension, and only at Commerce’s request. The draft bars every other US agency or department from suspending development, deployment, or internal use.

Who would evaluate the labs?

Independent evaluators registered with Commerce would periodically evaluate labs and frontier models. Labs could pay evaluators, as long as the payment isn’t tied to the outcome.

Did Anthropic confirm the draft?

No. Anthropic and the Commerce Department did not respond to Reuters’ requests for comment.

Sources: Politico; Reuters.