An Anthropic AI safety bill, still in draft form, would let the federal government block a frontier model only when the danger is extreme: more than $1 billion in damages, or more than 50 deaths or serious injuries. Politico reported the draft on October 16, 2026, after Anthropic privately pitched it to members of Congress. The proposal would also preempt some state AI laws, which puts it at the center of the fight over who sets the rules for advanced AI.

Congress has no comprehensive federal AI law. That gap is why a bill written by a leading lab deserves close reading, even though it has not been formally introduced. The details decide who can pull a model offline, how fast a company must report danger, and which state rules survive.

What the Anthropic AI safety bill would let Washington do

Under the draft, the government could intervene only when a frontier model presents a foreseeable, imminent risk of more than $1 billion in damages or of killing or seriously injuring more than 50 people, according to Interesting Engineering. The reporting frames the bill as one that limits federal power to halt models, and that limit is the core of the proposal.

Only the Commerce Department could seek a block. It would have to ask a federal judge in the U.S. District Court for the District of Columbia. The draft reportedly bars other federal agencies from suspending a model, a detail that Tom's Hardware summarized as making Commerce effectively the only agency capable of stopping an AI model.

The time limits matter too. Each judicial order would last up to 45 days, according to a summary of the draft from MLQ. Read literally, a suspension that runs past 45 days would need new orders, which keeps every block in front of a judge. That is an inference from the term length, not a separately reported provision.

Why the $1 billion and 50-death thresholds drive the debate

The two numeric triggers are where critics focus. Critics argue the bar is too high for harms such as AI-powered fraud, which falls outside the three covered categories, a point raised in Mother Jones's coverage. Mother Jones's headline calls the threshold alarming.

Another critic called the leaked draft unacceptable in a TechBytes post. Those objections point in different directions. One says the bar is too high for everyday harms, while the other rejects the draft outright.

The thresholds also differ from another set of numbers in circulation. CellCog describes Anthropic's public framework as covering models trained above 10^25 FLOPs, from firms with more than $500 million in AI revenue or more than $1 billion in AI R&D spending. CellCog presents those figures as part of the public framework, not the draft bill, so readers should not merge the two sets of thresholds.

The language also has a lineage. TechSpot, summarizing Politico's reporting, said the draft's threshold language mirrors California's AI laws. That gives lawmakers a familiar template, and it invites comparison with the state laws the draft would partly preempt.

Which risks the draft covers and which it leaves out

The draft limits covered risks to three categories: creating or releasing biological weapons, a model escaping its developer's controls, and cyberattacks, according to Law Commentary.

The list is narrow. Many AI harms that lawmakers already worry about, including fraud, would never reach the government's emergency powers under this draft. That design reflects a focus on catastrophic scenarios, but it also means everyday harms would stay under other laws and under state rules that the bill would partly preempt.

Evaluators, reporting duties and the 72-hour clock

The draft also builds a framework around outside testing and disclosure. Based on MLQ's summary and TechSpot's reporting, the key provisions include the following:

  • Outside evaluators would register with Commerce and be engaged periodically.
  • Evaluators must refer qualifying risks to Commerce within 72 hours.
  • Labs may pay evaluators, as long as the pay does not depend on what the evaluators find.
  • Frontier developers would face transparency and incident-reporting duties, including a 24-hour notice for an imminent risk of death or serious injury.

The pay condition separates an evaluator's fee from its conclusions, which is the first question any outside testing regime has to answer.

Compare this with California. Anthropic's statement says SB 53 requires the largest developers to publish frontier safety protocols and report serious incidents to the state within 15 days. The federal draft's 24-hour window applies to a narrower trigger, imminent risk of death or serious injury, but it runs far faster than SB 53's clock. A company operating under both regimes would need to track two very different deadlines.

How the draft would reshape state AI rules

The preemption language is the piece most likely to change the politics. The draft would preempt some state laws covering public risk disclosure, third-party evaluation and incident reporting, according to Yahoo's coverage. It would not reach all state AI rules. The sources reviewed do not publish a list of which state laws would fall.

States are already active. The same week, eight state attorneys general sued OpenAI over ChatGPT's child-safety claims, according to AI Weekly. The federal government is moving in a related direction. The White House's December 2025 AI National Policy Framework order seeks to challenge and preempt state AI laws that conflict with federal policy, per White & Case. Congress has enacted only one AI-specific law so far, the TAKE IT DOWN Act, which covers non-consensual intimate imagery, including deepfakes.

Anthropic's position needs context. The company has long argued that a uniform federal approach beats a patchwork of state laws, and it backed California's SB 53 while saying Congress should act. A draft that preempts some state rules on disclosure, evaluation and reporting raises a fair question about how those positions fit together. The coverage reviewed does not show the company answering it.

What the draft still leaves unsettled

  • Formal introduction: Tech Insider reported that the draft had not been formally introduced in Congress, and none of the coverage reviewed says it has been.
  • Timing: Whether the bill moves during the lame-duck session is unknown.
  • Preemption scope: The exact list of preempted state rules has not been published.
  • Origins: Coverage calls the draft privately pitched or leaked, and the sources do not agree on how it reached the press.
  • Counterproposals: Reporting has not said whether other labs or lawmakers will offer alternatives.

An Anthropic spokesperson, who was not named in the coverage, said the draft reflects the company's Advanced AI Framework, which it has promoted to federal and state lawmakers since June. The spokesperson added: “These are sensible policies we believe Congress should act on.” Tom's Hardware carried that statement. Mother Jones also reported that Anthropic has scaled up its Capitol Hill lobbying, which adds context to the timing of the draft.

What readers and businesses should do now

No action is required yet. The draft is not law, and it could change before any formal introduction. A few checks will show whether it matters to your work:

  • Watch for formal introduction. Search congress.gov for the bill title once it is filed, and check whether the Commerce role and the three risk categories survive.
  • If your business depends on frontier models, ask your AI vendors how they would handle a government suspension or a serious incident notice.
  • Review your state's rules on AI disclosure, third-party evaluation and incident reporting, since those are the categories the draft would preempt.

What comes next for the federal AI safety fight

The next signal is the first formal text. Lawmakers will have to decide whether to keep the $1 billion and 50-casualty triggers, whether Commerce should hold the only block power, and how much of the state framework to displace. Federal action against specific AI products is already happening. DHS ordered federal civilian agencies to remove DeepSeek products, as USA 24 reported. Anthropic's public profile is shifting too, with its IPO valuation talk reported to have slipped to $1.5 trillion from $2 trillion. The debate is moving from whether Washington needs an emergency brake to who holds it, how often it can be pulled, and which state rules get swept aside along the way.