Anthropic has confirmed it is circulating a draft federal bill among US lawmakers that would regulate frontier AI on the company’s own terms, according to an AFP report carried by Nigeria’s The Punch and Reuters wire copy published Friday. The draft — first reported by Politico on 16 October, citing three people familiar with the private pitch — would let Washington block a frontier model’s release or deployment only in three enumerated catastrophe scenarios, and would override tougher state AI laws in the process.
🔍 THE BOTTOM LINE
The company that built its brand on wanting to be regulated has now written the regulation — and the draft’s thresholds are set so high that essentially every documented AI safety incident of the past two years would fall outside its reach. A federal judge could halt a model only for imminent risks of killing or seriously injuring more than 50 people, or more than $1 billion in damages. It is a safety framework with the safety mostly designed out of it, and it arrives with a preemption clause that would strip US states of the authority to do more.
What the Draft Actually Says
The bill’s enforcement machinery — reported in Reuters’ itemised summary — works like this. Independent evaluators would register with the Commerce Department and periodically assess frontier labs and their models; labs could pay these evaluators, but not tied to the outcome of the evaluation. If evaluators find an “imminent enumerated risk,” they must report it to Commerce within 72 hours. Commerce could then ask a judge — specifically a federal judge for the District of Columbia — to suspend a model. No other department, agency, or officer could order a halt.
The three enumerated risk categories, per Tech Insider’s detailed read of the reporting, are narrow: a model contributing to the creation or release of a biological weapon, a model enabling a large-scale cyberattack, or a model evading its own developer’s safety controls. Everything else — biased output, hallucination, manipulative behaviour, underperformance — sits outside the bill entirely.
Beyond the emergency brake, the draft imposes standing obligations on every frontier developer: formal safety frameworks, published transparency reports per model, pre- and post-release testing against the enumerated risks, incident reporting to Commerce, and whistleblower protections for employees who raise safety concerns. An Anthropic spokesperson told AFP the proposal “grants the federal government the ability to block or deter dangerous deployments of the most powerful models” and “requires frontier developers to test the most powerful models, be transparent to the public about their findings, submit them to independent evaluation, and maintain a robust security program.”
The Part That Should Worry State Capitals
Buried in the structure is the clause with the widest consequences: by making Commerce the sole enforcement channel, the draft would functionally override the state AI laws passed since 2025. Anthropic itself wrote in June that “absent a strong federal law, states should retain full authority to legislate” — the new draft, as AFP reports, “would also override certain state laws, such as those in force in California.”
That is the tell. A company can simultaneously argue that 50 different state regimes are unworkable and push for a federal standard whose thresholds rarely bind. Anthropic’s draft does both at once. State officials who spent political capital on their own AI oversight regimes — California after SB 1047, Minnesota with its $500K-per-violation nudification ban, eight states currently suing over ChatGPT child-safety design — would see their standards displaced by a Washington that can act only above a $1 billion injury line.
Why It Mattered That Anthropic Moved First
The political context is not subtle. US policy under President Trump has leaned on voluntary pledges rather than statute; Congress, heavily lobbied and politically split, has passed no federal AI law and is out of session until after next month’s midterms. In that vacuum, whoever drafts first frames first. Anthropic — which called for a global AI pause in September and was forced by Washington in June to suspend its most advanced models for nearly three weeks — now supplies the reference text. A spokesperson said the bill “represents the proposals outlined in Anthropic’s Advanced AI Framework, which we have been advocating for with federal and state lawmakers since June,” per Politico’s reporting.
Reception among policy professionals is split in an instructive way. Charlie Bullock of the Institute for Law & AI called the draft “pretty reasonable” and “more than other frontier labs have done,” while still rating it “weaker” than the House’s competing FRONTIER Act, noted on X. Lawmakers appear split along the same seam: Senator Amy Klobuchar has pushed for verified testing of models, while Senator Ted Cruz captured the other camp’s instinct — “This is scary stuff, but we’re also not going to be able to stick our head in the sand.” The draft sits in the gap between them: real obligations on paper, a government trigger reserved for catastrophe.
The Conflict Nobody Is Naming
There is a structural conflict of interest embedded in this story that deserves plainer language than the coverage has given it. Anthropic is simultaneously the regulated party, the author of the proposed regulatory text, and — through its evaluator-funding mechanism — a payer of its own assessors, with only a handshake rule (no outcome-tied payment) separating the checker from the checked.
The pattern is not new to 2026. The company already publishes model safety assessments voluntarily; codifying your own existing practices as a legal floor is the classic incumbent move, because every competitor without the same infrastructure inherits a compliance cost you already absorbed. What is new is the speed: from “please regulate us” to “here is the bill” in under a year, timed to a Congress with nothing else on the AI docket.
New Zealand is watching this from an awkward distance. Our own framework bet — personal responsibility over hard frontier rules — looks small next to a superpower drafting its own AI statute with the industry’s pen. But the lesson travels: the country that writes the template first exports it. If Washington enacts something shaped like this draft, expect the next round of “harmonisation” pressure on every small jurisdiction with AI rules of its own — Australia summoned the same company’s CEO to explain rogue agent incidents, and frameworks collide at exactly that level.
❓ FAQ
Has Anthropic’s bill been introduced in Congress? No. As of publication the draft has not been formally introduced; it was pitched privately to Capitol Hill staff, which is typically an early move to gauge interest and find a sponsor. Congress is out of session until after the November midterms.
When could the government actually stop a model under this draft? Only if evaluators report an imminent risk in one of three categories — bioweapon contribution, large-scale cyberattack, or model evasion of developer controls — with foreseeable harms above $1 billion or 50-plus people killed or seriously injured, and only via a DC federal judge after a Commerce request.
Would state AI laws survive it? The draft appears designed to preempt them, with Commerce as the sole enforcement channel — a direct shift from Anthropic’s June position that states should retain full authority absent a strong federal law.
Why does Anthropic get to write the bill? Nobody granted it formally. The company moved first with detailed legislative text in a policy vacuum, and first drafts frame everything after them. Rivals like OpenAI have engaged statehouses separately — OpenAI backed a narrower Massachusetts bill earlier this year — but no other lab has produced a comparable federal draft.
🔍 THE BOTTOM LINE
Watch the thresholds, not the transparency chapters. Every serious provision in this draft — evaluators, reporting clocks, whistleblower shields — funnels into a trigger set so high it has never been met by any documented AI incident, and the same draft takes state regulators off the board. If Congress writes anything this year, the fight will be over those two numbers: $1 billion and 50 people. They are not technical details; they are the whole debate, and Anthropic just opened it holding the pen.
📰 Sources
- Politico — “Anthropic quietly pitches its own AI safety bill to lawmakers” (16 October 2026)
- Reuters via WHBL — itemised wire summary of the draft bill’s mechanics (16 October 2026)
- AFP via The Punch — “Anthropic proposes AI safety bill to US lawmakers” (17 October 2026)
- Tech Insider — “Anthropic Pitches a $1B AI Safety Bill to Lawmakers” (17 October 2026)
- Dnyuz (Politico syndication) — same reporting (16 October 2026)
