Anthropic Pushes Federal AI Safety Bill that Limits Government Intervention to Catastrophic Thresholds

Anthropic has drafted legislation allowing U.S. federal agencies to seek a court order before blocking the development, deployment, or release of an advanced AI model. The bill defines a “catastrophic risk” as a foreseeable, imminent scenario that could cause death or serious injury to more than 50 people or over $1 billion in property damage from a single incident. An independent evaluator registered with the Department of Commerce must notify the department within 72 hours if such a risk is identified. The department would then petition a federal judge in Washington, D.C. to halt the model’s release. Agencies cannot stop a model on their own; the final decision rests with the judiciary.

Anthropic Draft Bill Would Let Judge Halt AI Model Release

The draft mirrors the numerical thresholds used in California’s SB 53, which requires AI developers to assess and disclose catastrophic risks. Anthropic has promoted its “Advanced AI Framework” to lawmakers since June, positioning the draft as a national standard that builds on state-level guidance.

Rogue AI Must Threaten 50 Lives Before Being Shut Down, Anthropic Suggests

• Anthony Cimino – Head of federal affairs at Anthropic; noted the thresholds are drawn from California law and said the bill would not shield companies from liability.

Politico

• Zvi Mowshowitz – Writer covering AI policy; called the proposed thresholds “absurdly high.”

Rogue AI Must Threaten 50 Lives Before Being Shut Down, Anthropic Suggests

• Nathan Calvin, Nat Purser, and Nate Soares – AI safety experts who warned the bill could weaken safeguards and limit oversight.

Rogue AI Must Threaten 50 Lives Before Being Shut Down, Anthropic Suggests

• Charlie Bullock – Analyst at LawAI; warned the draft lacks standards for auditors and could encourage the cheapest, least burdensome reviews.

Anthropic Proposes AI Safety Bill Letting U.S. Agencies Block Models Over $1 Billion Damages Or 50 Deaths: how 16 outlets framed it

• Thresholds: > 50 people at risk of death or serious injury; > $1 billion in property damage.

Anthropic Draft Bill Would Let Judge Halt AI Model Release

• Evaluation timeline: Independent evaluators must report qualifying risks within 72 hours.

Anthropic Proposes AI Safety Bill Letting U.S. Agencies Block Models Over $1 Billion Damages Or 50 Deaths: how 16 outlets framed it

• Judicial venue: Court order sought from a federal judge in Washington, D.C.

Anthropic's AI Safety Bill Hides Judge-Only Block Clause

Anthropic’s spokesperson told *Politico* the proposal expands the company’s Advanced AI Framework and aims to create a “strong national standard.”

Politico

AI safety researchers Calvin, Purser, and Soares argue the bill would restrict federal oversight and undermine existing safety mechanisms. Bullock highlighted the absence of licensing or standards for independent auditors, warning that auditors might compete by offering the “least burdensome and cheapest reviews,” potentially compromising risk assessment.

Anthropic Proposes AI Safety Bill Letting U.S. Agencies Block Models Over $1 Billion Damages Or 50 Deaths: how 16 outlets framed it

Anthropic has not set a legislative timetable. The proposal remains pending congressional consideration and may be amended in response to industry and expert feedback.

Politico