Anthropic’s draft AI safety bill would allow a frontier model’s release to be halted only by a federal judge in Washington, D.C., and only after registered evaluators find a foreseeable, imminent risk of more than 50 deaths or more than $1 billion in property damage. The draft, first reported by Politico on October 16, 2026, expressly bars any federal agency from stopping a model on its own authority.
The Anthropic AI Safety Bill Sets a 50-Death, $1 Billion Trigger
The draft defines the risk that can stop a release with two numbers: more than 50 people at risk of dying, or more than $1 billion in property damage. According to Politico’s reporting, as summarized by TechCrunch senior writer Lucas Ropek on October 20, the risk must also come from one of three sources: the creation or release of a biological weapon, a model evading its developer’s controls, or a cyberattack. Those figures match the ones California uses in Senate Bill 53 to define catastrophic risk. Anthropic’s own Advanced AI Framework, published in June 2026, explicitly recommends a catastrophic risk definition in line with SB 53.
Registered Evaluators Would Have 72 Hours to Alert the Commerce Department
Under the draft, the judgment that a model is dangerous belongs to independent evaluators registered with the U.S. Commerce Department. Politico reported that the labs themselves could pay those evaluators, provided the payment is not contingent on the results. Once an evaluator concludes that a model poses an imminent catastrophic risk, it has 72 hours to tell the Commerce Department. The department then has to ask a judge in Washington, D.C., to stop the release. An Anthropic spokesperson confirmed the draft to Politico and said it represents the proposals in the company’s Advanced AI Framework, which Anthropic has been pressing on federal and state lawmakers since June.
Anthropic’s June Framework Offered Several Enforcement Options, and the Draft Picks the Narrowest
The 19-page framework set out a menu of ways to block or deter unsafe models and said policymakers could mix and match them. One option let an agency impose fines or deployment bans after reviewing system cards, risk reports and evaluator findings. The safeguards it listed alongside that option included requiring the agency to go to court instead of imposing remedies directly, and limiting the agency to specific listed violations instead of its own assessment of risk. The draft bill, as Politico described it, adopts the court route and removes agency discretion entirely. The framework also proposed obligations the reporting on the draft does not detail: coverage for developers training models above 10^25 FLOP with more than $500 million in annual AI revenue or more than $1 billion a year in AI research spending, reporting of critical safety incidents within 15 days, and published risk reports at least every six months.
Evaluator Independence Is the Part the Whole Mechanism Rests On
Because no agency can act alone, every block in the draft starts with an evaluator’s finding, which makes the evaluator’s independence the load-bearing piece. Ropek wrote that it is unclear how the proposal would stop lab payments from shaping evaluations. Anthropic’s framework names the same problem. It warns of “evaluator shopping,” where developers seek the evaluator who asks the least of them, and suggests government ratings of evaluators and random assignment in high-stakes cases. It also floats government funding or a pooled industry fund so evaluators can stay financially independent of any one developer, plus cooling-off periods for staff. Politico’s description of the draft retains direct lab payment with a results-contingency ban, which is a lighter safeguard than the pooled funding the framework discussed.
Why This Matters for AI Governance and Procurement
The draft has no force of law, but it shows where one frontier lab wants the checks on its models to sit, and that location is the outside evaluator. Teams buying frontier models can apply the same test in vendor reviews today. Ask which external evaluator examined the model you are licensing, whether that evaluator published its findings, who paid for the work, and whether the payment depended on the outcome. Ask for the most recent system card and risk report, and check whether the evaluator had access to an unredacted version, which is what Anthropic’s framework says evaluators should receive. If a vendor cannot answer those questions, the evaluation it cites carries little weight in a procurement file. If the draft or something like it passes, those same answers become a regulatory record, and the teams that already ask for them will have less to change.
FAQ
What does Anthropic’s draft AI safety bill propose?
It would let a federal judge halt the release of an AI model when registered independent evaluators find a foreseeable, imminent risk of more than 50 deaths or more than $1 billion in property damage, according to Politico’s reporting.
What risks would trigger a block under the Anthropic bill?
The risk has to come from the creation or release of a biological weapon, a model evading its developer’s controls, or a cyberattack.
Could a federal agency stop a model on its own under the draft?
No. The draft forbids any federal agency from halting a model’s distribution at its own discretion; the Commerce Department would have to ask a judge in Washington, D.C.
Who would pay the independent evaluators?
The AI labs could pay them, as long as payment does not depend on the evaluation’s results, Politico reported.
How is the bill related to Anthropic’s Advanced AI Framework?
An Anthropic spokesperson said the bill represents the proposals in the June 2026 framework, which the company has been advocating to federal and state lawmakers since June.
Is the Anthropic AI safety bill law?
No. It is a draft Anthropic has been circulating in the hope that Congress takes it up.