Anthropic's draft frontier AI bill: block a model only above $1 billion or 50 deaths

Anthropic has privately pitched members of Congress a draft federal frontier AI bill under which the government could block a model only for an imminent risk of more than $1 billion in damage or more than 50 deaths or serious injuries. Politico obtained the unreleased draft and reported it on October 16, 2026. It is not public, has no reported sponsor, and Anthropic says its preemption language will be revised.

What the draft would do

Per Techstrong.ai's October 17 account of the Politico reporting, federal authorities could stop a model's release or deployment only if a threat is both foreseeable and imminent and involves one of three scenarios: creating or releasing biological weapons, a massive-scale cyberattack, or an AI model evading its developer's internal safety controls. The damage threshold, more than $1 billion or more than 50 people killed or seriously injured, is described as reflecting language from California's SB 53. Our comparison of New York's RAISE Act and SB 53 covers that state baseline.

The mechanism is deliberately narrow. Registered independent evaluators, funded by labs but not paid according to their findings, would run regular reviews of frontier models. MLQ, summarizing a circulated copy dated September 29, says those evaluators would be engaged every 180 days and registered with the Commerce Department. If an evaluator finds an imminent enumerated risk, it must notify Commerce within 72 hours. Only Commerce could then ask a federal judge in the US District Court for the District of Columbia for an injunction. The bill bars other federal departments from suspending a model's development or deployment on their own.

MLQ adds details Techstrong omits. Each court order could last up to 45 days, and the government would have to show no less restrictive measure would work. Critical incident reports would be due within 15 days, with 24 hours' notice when an incident creates an imminent risk of death or serious injury.

Who it would cover

Larger developers, defined as those above $500 million in annual AI revenue or $1 billion in annual AI spending, would face added duties, including safety frameworks and annual compliance certification, per MLQ. Other provisions described are transparency reports with new general-access deployments, whistleblower protections and incident reporting.

An Anthropic spokesperson confirmed the outreach and described the draft as reflecting its Advanced AI Framework. The spokesperson called it "sensible policies we believe Congress should act on," per Techstrong.ai.

The fight over preemption

The draft would preempt certain state requirements on public disclosures, third-party evaluations and incident reporting, while preserving state tort, child-safety and procurement rules. That is the contested part. Techstrong.ai reports that tech industry groups favor a single national framework, while progressive lawmakers oppose preemption, pointing to safety laws already in place in California, Illinois and New York. MLQ says Anthropic's federal-affairs lead said the language would be revised to give states a role in evaluator standards and shutdown rules, so the text may change.

The Center for Democracy and Technology has warned that a catastrophic-risk focus can miss civil-rights and consumer harms, per MLQ. That critique matters here because a bill that triggers only above $1 billion or 50 deaths addresses a small set of extreme events, not the everyday harms that state laws also target.

Where it sits in a crowded field

Techstrong.ai lists the competing efforts. In the House, Reps. Lori Trahan and Jay Obernolte introduced the FRONTIER Act. In the Senate, bipartisan talks led by Majority Leader John Thune, Amy Klobuchar and Ted Cruz continue amid pushback from lawmakers who want stronger government review powers. Sen. Bernie Sanders has pushed for stricter measures, including temporary pauses on advanced AI development, and we track his bill in our explainer on the Ban Artificial Superintelligence Act. Sen. Mark Warner's bill, covered in our S. 5576 piece, takes the opposite tack with pre-release access.

Techstrong.ai also notes that major labs disclosed incidents over the summer in which AI agents probed or hacked online platforms, but none reportedly meets the damage threshold. That is the sharpest question the draft raises. A high bar for blocking a model is also a bar the summer's incidents did not clear.

The case for, and against

For the draft: it names concrete triggers, forces independent evaluation, creates a single point of authority, and limits the risk of a patchwork of state rules. Anthropic has already moved in this direction, as its arrangement giving METR employee-like access shows.

Against: the thresholds are high enough that a serious but sub-threshold event could go unaddressed. Evaluators are lab-funded. And a lab writing the rules for its own industry is a conflict, whatever the quality of the text. MLQ also notes that talks among Anthropic, OpenAI and Google DeepMind on third-party evaluators have raised antitrust concerns.

What to read next in the filing

If a sponsor does introduce text, four lines will tell you whether the public description holds. The definition of covered developer, since the reported $500 million revenue or $1 billion spending test decides who is bound. The evaluator registration rules, since independence is the whole safeguard. The preemption section, since it decides whether California's SB 53 and New York's RAISE Act survive intact. And the remedy: a 45-day order that Commerce must justify against less restrictive options is a very different tool from an open-ended suspension.

Our take

This is a lobbying document, and it should be read as one. It is also the most specific text any lab has put in front of Congress, which makes it useful: you can argue with numbers. Watch three things. Whether the preemption section survives. Whether a sponsor attaches their name to it. And whether anyone proposes a lower threshold than 50 deaths. Until a sponsor does, nothing here binds anyone. Neither Politico's original nor the draft itself was available to us, so every detail above is as reported by the two outlets named.