Anthropic Quietly Pitches Congress Its Own AI Safety Bill
Anthropic has written its own federal AI safety bill and has been quietly showing it to lawmakers in Washington. Politico first reported the draft on October 16, 2026, and Anthropic has since confirmed it is circulating the proposal. The company says it reflects the policies it has been pushing lawmakers to adopt since June, under its own Advanced AI Framework.
The draft is not a law. No member of Congress has introduced it, and a document a company hands to lawmakers can change a lot, or go nowhere, before it ever gets a vote. But it is a rare case of a frontier AI lab writing the actual legal text it wants passed, rather than just lobbying for general principles. Here is what the draft would actually do, and why it is already drawing pushback.
⚡ Quick facts
- What happened: Politico reported Oct 16, 2026 that Anthropic is privately circulating a draft federal AI safety bill to lawmakers
- Not yet introduced: No member of Congress has introduced the draft or signed on as a sponsor
- Trigger threshold: A foreseeable, imminent risk of 50+ deaths/serious injuries or $1 billion+ in damages from a bioweapon, a rogue model, or a cyberattack
- Who can act: The Commerce Department, only after a federal judge in Washington, D.C. approves blocking the model
- Preemption fight: The draft would override parts of state AI disclosure laws, including California's SB 53, which Anthropic itself backed in 2025
What Anthropic is proposing
The draft would require frontier AI developers to test their most powerful models before release, publish transparency reports on what those tests found, protect employees who report safety concerns, and disclose critical safety incidents. None of that is unusual; versions of these obligations already exist in state laws and in other federal bills working through Congress.
What makes Anthropic's draft different is the emergency power it creates. It would let the federal government step in and block a model's release, or its continued use, but only in a narrow set of circumstances defined in the bill itself.
When the government could actually step in
The bar is high by design. The draft limits federal intervention to a foreseeable, imminent risk of killing or seriously injuring more than 50 people, or causing more than $1 billion in damages. It also narrows the covered scenarios to three: a model helping someone build a biological weapon, a model escaping its developer's controls and acting on its own, or a model enabling a major cyberattack. Risks outside those three categories, however serious, would not qualify.
Who gets to pull the trigger
Under the draft, an independent evaluator who spots one of those threats would have to report it to the Commerce Department within 72 hours. Commerce would then bring the evidence before a federal judge in Washington, D.C., and the judge, not Commerce, would decide whether to approve an order blocking the model. No other federal agency could act on its own to halt a release. That two-step structure, an evaluator flagging the risk and a court approving the response, is meant to stop any single official from shutting down a model on their own judgment.
The fight over state AI laws
The most contested part of the draft is what it would take away from states. Reporting says it would preempt state rules on public disclosures, third-party evaluations and incident reporting, the same categories covered by California's SB 53, the Transparency in Frontier AI Act. Anthropic was the first major AI company to publicly back SB 53 in 2025, before it took effect in January 2026, making this draft a reversal of sorts: the company that helped build the case for state-level transparency rules is now pushing a federal bill that would override them.
California has also been moving on a second, separate track. Governor Newsom signed SB 813 and AB 1405 in September 2026, creating a state AI Safety Commission and an auditor registry. That law is not the one Anthropic's draft targets, but it shows how much state-level AI rulemaking the federal bill would need to work around if it preempts broadly.
Why critics call the bar too high
Nathan Calvin, general counsel at Encode AI and one of the people who helped shape SB 53, has said the draft has some sensible elements but that passing it, in his view, would leave the public less protected than current law. His objection centers on the combination of a high casualty-and-damage threshold with broad preemption: if the federal trigger rarely fires and it still cancels out stronger state rules, the net effect is less oversight, not more.
Charlie Bullock of the Institute for Law & AI has taken a more generous view, calling the draft "pretty reasonable" while still rating it weaker than the FRONTIER Act, a rival bill introduced in the House in July by Reps. Jay Obernolte and Lori Trahan. His specific criticisms point to the draft's rulemaking process and its emergency shutdown authority as the places it falls short.
What happens next
Congress has several competing AI bills in play right now, including the FRONTIER Act, Senator Elissa Slotkin's AI Guardrails Act of 2026, and a national AI safety framework from Senate Majority Leader John Thune that has been pushed back until after the chamber's recess. Anthropic's draft is not formally attached to any of them, and with Congress heading toward the midterm elections, none of these proposals is likely to move quickly. For now, the draft is a marker of where Anthropic wants the eventual federal rules to land, not a bill on its way to a vote.
Frequently asked questions
What is Anthropic's draft AI safety bill?
It is a federal bill Anthropic wrote and has been privately circulating to US lawmakers. Politico first reported it on October 16, 2026. It would let the government block a frontier AI model only after an independent evaluator flags a foreseeable, imminent risk of mass casualties or billion-dollar damage.
Has Anthropic introduced the bill in Congress?
No. As of this report, no member of Congress had introduced the draft or signed on as a sponsor. A circulated draft can change substantially before, or if, it is ever introduced.
What would trigger a federal pause under the draft?
A foreseeable, imminent risk of killing or seriously injuring more than 50 people, or causing more than $1 billion in damages, arising from a biological weapon, an AI model escaping its developer's controls, or a major cyberattack.
Who would have the power to block an AI model?
An independent evaluator would report an imminent threat to the Commerce Department within 72 hours. Commerce would then have to bring that evidence before a federal judge in Washington, D.C., who would need to approve any order blocking the model. No other federal agency could act on its own.
Would the bill override California's AI laws?
Reporting says the draft would preempt parts of state AI laws covering public disclosures, third-party evaluations and incident reporting, including California's SB 53, the Transparency in Frontier AI Act that Anthropic itself publicly backed in 2025.
Why do critics say the bar for action is too high?
Encode AI's Nathan Calvin has said the bill has sensible parts but that, in his view, passing it would leave the public less protected than existing state law, pointing to its preemption language and the high threshold required before Washington can act.
Comments