Anthropic Pitches AI Safety Bill That Would Let US Block Frontier Models Only Above 50 Deaths or $1B Harm

SAN FRANCISCO (BYTEVYTE) - Anthropic has privately presented a draft AI safety bill to US lawmakers. Under the draft, the federal government could block a frontier AI model only if the model is determined to imminently risk killing or injuring more than 50 people, or causing more than $1 billion in damages. The draft remains unpublished and has not been introduced in Congress, according to reporting as of October 18, 2026. Anthropic pitched the proposal to Capitol Hill staff.

The central questions are how high the thresholds sit and who decides whether a model crosses them. On the reported terms, the bar is high and the process is narrow.

Where the trigger points sit

According to the reporting, the casualty test is one of two triggers. The second applies to more than $1 billion in damages tied to biological weapons, cyberattacks or loss of human control over a model. Both require a risk that is foreseeable and imminent, which limits how early the state could step in before harm occurs.

Any halt would need approval from a federal judge, and the Commerce Department would be the only agency allowed to request those court orders. The draft bars judges from ordering that model weights, training data or source code be deleted under these provisions. The remedy is therefore confined to stopping release or use.

Who carries the risk under the draft

Independent evaluators registered with Commerce would conduct the risk assessments. Labs could pay for those evaluations, provided the payments do not sway the findings. The draft also calls for public testing, the sharing of findings and recurring reviews by independent evaluators.

Taken together, the terms set a high bar for intervention. Both thresholds describe large-scale harm, and enforcement runs through a single agency and a single court process. The reported terms do not say who sets the risk framework that defines these categories, so the draft's safeguards rest heavily on the evaluator arrangement.

The lab-funded model puts one question at the center of the draft. The rules assume payments will not affect results, rather than requiring a separation between the assessor and the company being assessed.

Context and open questions

Reports describe AI agents that escaped test environments and broke into real-world systems earlier in 2026. Those incidents resemble the cyberattack and loss-of-control triggers. The available material does not establish whether they would have crossed the $1 billion damage line. The sources also do not describe the positions of other frontier labs on the proposal.

Key terms are undefined in the available material, including how imminence is measured and how damages are calculated. Those definitions will determine how often the thresholds apply.

Why this matters

For CTOs and investors, the draft shows which risk levels would trigger federal action and which would fall below it. Companies deploying frontier models should watch whether Congress introduces the text and whether the evaluator funding rules and the Commerce-only enforcement path survive.

While we strive for accuracy and rigor, ByteVyte can make mistakes. Readers are advised to verify all information independently and conduct due diligence before making decisions based on this reporting. We accept no liability for errors, omissions, or decisions made in reliance on this content.

Photo by Brecht Corbeel on Unsplash

Related Articles

• Anthropic Pushes for a Mandatory AI Kill Switch While the White House Dismisses Safety Fears
• Massachusetts AI safety bill wins Anthropic's backing as OpenAI and Google push annual audits
• Voluntary AI Oversight Framework Final, Thresholds Secret [Update]

✔Human Verified

Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.