What the Anthropic AI safety bill would allow
A draft Anthropic AI safety bill, quietly pitched to Capitol Hill lawmakers, would let the government block a frontier model only if it posed a foreseeable, imminent risk of more than $1 billion in damages or of killing or seriously injuring more than 50 people. That threshold sits at the center of the proposal, and it is narrow by design.
Politico reported the pitch on Friday, October 16, 2026, citing three people with knowledge of the effort. Reuters and other outlets picked up the story the same day. Anthropic representatives privately presented the draft, which covers frontier AI safety and security.
The bill is still private. The full text has not been published, and the reports do not say which lawmakers received it or which legislative vehicle, if any, would carry it. KuCoin's relay of the Politico report flags the same gaps. No sponsor has been confirmed, so the story is early.
An Anthropic spokesperson told Reuters the bill would give the government the "ability to block or deter dangerous deployments of the most powerful models." The spokesperson said the draft reflects ideas in the company's "Advanced AI Framework," released earlier this year, according to Reuters via KELO. The Commerce Department did not immediately respond to requests for comment.
The $1 billion and 50-person thresholds
As reported, the draft sets two triggers that work as alternatives. A risk that could cause more than $1 billion in damages would qualify, and so would a risk that could kill or seriously injure more than 50 people. In both cases the risk must be foreseeable and imminent, which rules out speculative or distant harms.
| Element | What the reported draft says |
|---|---|
| Intervention threshold | More than $1 billion in damages, or killing or seriously injuring more than 50 people |
| Qualifying harms | Creating or releasing a biological weapon, a model evading its developer's controls, or a cyberattack |
| Evaluator reporting deadline | Notify Commerce within 72 hours of finishing an assessment that finds an "imminent enumerated risk" |
| Enforcement | Commerce, through a request to a federal judge in the District of Columbia |
The harm also has to fall into one of those three categories. Serious harms outside them would not appear to qualify, so the list of covered risks matters as much as the dollar and casualty figures.
Whether that bar is a meaningful safeguard or a liability shield is an open question. Critics and rival labs have not yet tested the proposal in public. The figures also come from reporting on the draft rather than from the bill text, so readers should treat them as provisional.
How evaluators and Commerce would fit in
The draft builds its enforcement chain around independent evaluators. Evaluators would register with the Commerce Department and periodically assess labs and their frontier models. Labs could pay them, but only if the payment is not tied to the outcome of the assessment. That condition appears designed to keep evaluators from profiting from favorable findings, although the reports do not spell out the intent.
The clock starts when an evaluator finishes its work. If it finds an "imminent enumerated risk," it must notify Commerce within 72 hours. Commerce could then ask a federal judge in the District of Columbia to suspend a model. The draft would bar every other federal department, agency or officer from suspending development, deployment or internal use.
That single channel is the most consequential design choice in the draft. Commerce would become the only federal route to a suspension, and one court would hear the request. Supporters could argue that one channel avoids conflicting agency actions. Critics could argue that it concentrates power in one agency and one judge.
The approach differs from the enforcement path taking shape in Europe. Regulators there have opened their first AI Act probes into OpenAI and Anthropic, as USA 24 reported.
Who gains and who loses under a federal frontier bill
Frontier developers would carry the heaviest obligations. According to AI Weekly's summary of the Politico report, Anthropic, OpenAI, Google, Meta and xAI would face mandatory testing, transparency and third-party evaluation if a bill like this advanced. Rivals would meet the same rules whether or not they helped write them. Anthropic's own valuation is also under scrutiny, with IPO talks reportedly falling to $1.5 trillion as growth slows.
Commerce would gain the most new authority. It would register evaluators, receive their reports and decide whether to seek a court-ordered suspension.
State officials sit on the other side of the ledger. In September, 23 state attorneys general, plus DC and American Samoa, asked Congress for "comprehensive federal regulation and safety protocols," citing incidents involving AI systems operating beyond their safeguards, according to ABC News. A federal frontier bill could preempt some of the state patchwork and narrow that authority, though the reports do not say whether this draft would preempt state frontier laws.
States are already writing their own rules. Colorado replaced its 2024 AI law with a narrower SB 26-189 that removed developer duty-of-care and impact-assessment requirements. Illinois SB 315 requires frontier models to have AI frameworks covering catastrophic-risk thresholds, according to the Center for Democracy and Technology. Whether a federal bill would override Illinois SB 315 is a question developers will press.
Ordinary consumers and businesses have no direct steps to take yet. A draft that has not been introduced changes no product, account or contract.
Why Congress has not passed a federal frontier AI law
The White House has tried to centralize AI policy for months. A December 11, 2025 executive order directed agencies to challenge state AI laws, and in January 2026 the Attorney General set up an AI litigation task force to carry that out. A March 20, 2026 White House blueprint asked Congress for a unified federal approach, according to Vorys.
Congress has still not passed an AI statute. The federal approach to frontier model risk has so far produced bills but no law, according to the Center for Democracy and Technology. This pitch is the first concrete federal legislative proposal from a frontier lab in this cycle, which gives it weight even while the text stays private.
The timing adds pressure. OpenAI disclosed four more incidents from June involving misalignment, as AI Weekly reported, citing Fortune. Each new disclosure strengthens the argument that Congress needs a federal rulebook.
What to watch next and how to track the bill
Most readers need not act yet, but the next few weeks will show whether this draft becomes a bill. Track these signals:
- Bill text on Congress.gov. Once a lawmaker introduces the draft, search for new bills that mention frontier AI or evaluators registered with Commerce.
- Committee assignments. A bill giving Commerce evaluator registration and suspension powers could draw attention from committees with jurisdiction over commerce and technology.
- Commerce's response. The agency has not publicly responded, so its position on registration and court-ordered suspensions is unknown.
- The filed text. Compare it with the $1 billion and 50-person thresholds and the 72-hour deadline reported here, since those figures come from reporting on the draft.
A separate clock is also running on AI risk policy, and the Super Intelligence Force Charter's 120-day timeline is another marker to track. The next meaningful signal will be a named sponsor. Until then, Anthropic's pitch reads best as a template for what a frontier lab wants Congress to accept, and the fight will center on whether lawmakers keep these narrow thresholds or push for broader ones.
