Anthropic drafts bill letting courts halt risky AI

By VK

Published 21 Oct 2026, 10:56 UTC

Reporting from TechCrunch

Anthropic confirmed a draft federal bill that, per Politico, would let a court halt a model release over a risk of 50+ deaths or $1 billion in damage.

What it means for founders

Most startups sit outside the scope. If the bill tracks the framework, duties land only on labs with $500 million in AI revenue or $1 billion in yearly research spend. Teams building on models would feel it indirectly, through release timing and provider risk reports.
A court halt is a new supplier risk. A model you plan to adopt could be held back by a judge after an evaluator's warning. Keep a second provider tested so a frozen release does not freeze your roadmap.
Evaluation could become a paid market. If labs pay registered evaluators, as the reported draft allows, that creates demand for independent testing in biosecurity, cyber and agent control, an opening for security startups with genuine depth.
Watch Congress and the states. The signals are a named sponsor, a hearing, and whether federal rules try to override state laws like SB 53, which the framework opposes unless federal rules match its own toughest proposals.

The story

Anthropic has been circulating draft federal legislation that would let a court stop an AI model from being released when independent testers judge that it poses an imminent catastrophic risk, Politico reported. A company spokesperson confirmed the draft to the outlet, saying the bill puts into legislative form the Advanced AI Framework Anthropic has pitched to federal and state lawmakers since June. TechCrunch summarized the reporting on Tuesday, October 20.

How the draft bill would work

According to Politico's account, a release could be blocked when the danger is foreseeable and imminent, putting more than 50 lives at risk or threatening over $1 billion in property damage. The harm must also trace to one of three sources: a biological weapon, a cyberattack, or a model escaping its developer's controls. Those figures match the catastrophic risk bar in California's SB 53.

Per Politico, the checks would fall to independent evaluators registered with the Commerce Department. Labs could pay them, provided the fee does not hinge on the findings. An evaluator that spots an imminent catastrophic risk would have 72 hours to alert Commerce, which would then have to ask a judge in Washington, D.C., to halt the release. No federal agency could stop a model on its own say, Politico reported.

How it compares with the June framework

The published framework goes well beyond the reported bill. It would cover developers that train above a set compute threshold and also earn more than $500 million a year from AI or spend over $1 billion a year on AI research. They would publish safety frameworks, system cards and twice-yearly risk reports, and report serious incidents within 15 days. It names four risk areas, adding automated research and development to the three in the draft.

On blocking unsafe models, the framework offered a menu, from agency review to court enforcement, and cautioned against overly broad regulatory power. The reported draft appears to lean on the court enforcement option: Commerce could not halt a release itself and would have to go to a judge. The framework also warns that labs may hunt for the most lenient evaluator, and floats government ratings and random assignment as fixes.

TechCrunch notes harms this large are some way from recent agent hacking incidents, such as OpenAI's test agents breaking into websites.

What we don't know yet

Whether any member of Congress will sponsor or introduce the bill.
What the full bill text says, including exactly which developers it covers.
How Commerce would register evaluators, and what would stop lab payments from shaping results.