Anthropic is pitching Washington lawmakers its own bill on the safety of powerful AI. A spokesperson confirmed to Politico and Reuters that the proposal draws on its Advanced AI Framework, published in June. The company points to model testing, public reporting of findings and independent evaluations as central provisions.
According to Politico, the government could seek to stop release or use only for foreseeable, imminent risks exceeding $1 billion in damage or more than 50 deaths or serious injuries. The reported categories concern biological weapons, cyberattacks and models evading their developers' controls.
Reuters also reports periodic evaluations by outside assessors registered with the Commerce Department, with a suspension proceeding through a federal court. These details come from reporting on the confidential draft; the reach of government intervention powers remains politically contested.
In his own post, Alex Bores argued that a proposal based on the June framework had not kept pace with recent developments. Zamaan Qureshi called for stronger internal auditor access, timely reporting and room for tougher state laws.
Charlie Bullock welcomed a frontier lab putting forward a concrete safety bill. He praised proposed continuing audits and information access, while judging its rulemaking and emergency shutdown powers weaker than those in the FRONTIER Act. That is his assessment of the proposal.
Nat Purser proposed specific changes: auditors should not be selected or paid by the labs they oversee, authorities should be able to compel corrective action, and minimum standards should evolve with capabilities. The dispute centers on how much independent control a federal law would actually provide.