Anthropic, the company behind the Claude models, has handed US lawmakers a confidential draft of a federal bill on frontier AI safety and security, according to Politico.
Politico reported the draft on October 16, 2026. The proposal would give the federal government limited authority to step in on the deployment of highly advanced AI models. That power would apply only when a model poses a foreseeable, imminent catastrophic risk.
What the draft bill would do
The draft sets out strict thresholds for when Washington could act. Intervention would be allowed where potential damages exceed $1 billion, or where AI use could cause the deaths or serious injuries of more than 50 people.
The bill ties those thresholds to three specific danger zones. These are biological weapons, a developer losing control of its own model, and cyberattacks.
Testing is the second pillar. The draft calls for independent testing, more transparency, and ongoing evaluation of advanced models.
Those independent evaluators would need to register with the Commerce Department. They would conduct regular assessments, and they could not be paid based on outcomes.
Under the proposal, emergent risks would have to be reported to the Commerce Department within 72 hours.
One agency, one judge
Only the Commerce Department would have the authority to ask a federal judge for an order halting a model’s deployment. Other agencies would be expressly barred from intervening.
Built on a California blueprint
The proposal mirrors aspects of California’s SB 53, which was signed into law in 2025. Anthropic appears to be using that state-level law as a template for a national version.
The company has also been pushing an “Advanced AI Framework” with lawmakers since June 2026. The draft bill is the most concrete form that advocacy has taken so far.
What this means for the AI industry
Mandatory independent evaluations and transparency requirements would mean new spending for any company building frontier models. Firms would likely need dedicated safety teams, reporting systems, and relationships with registered evaluators.
The thresholds and risk categories suggest the bill is aimed at the most advanced systems, not the average chatbot startup.
The independent evaluation requirement could also create a new market. If the draft became law, registered third-party testers would become a required part of the frontier AI supply chain.
The $1 billion damage line and the 50-person casualty line are both high bars. That keeps the government’s emergency power focused on catastrophe, but it also means most AI harms would fall outside the bill’s intervention authority.
For now, the draft remains confidential and has not been formally introduced.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.