Anthropic has circulated a confidential draft bill to Capitol Hill lawmakers that would establish federal oversight of frontier AI models while restricting government intervention to specific catastrophic risk thresholds.
The proposed legislation, confirmed by an Anthropic spokesperson as an extension of the company's internal Advanced AI Framework, would create a statutory mechanism for evaluating frontier AI systems and preempt state-level safety regulations.
Intervention Thresholds and Catastrophic Risk Scopes
Under the draft proposal, federal regulators would operate under strict statutory constraints regarding when they can halt or restrict the deployment of an artificial intelligence model. The government could block a model's release or continued operation only if it demonstrates a foreseeable, imminent risk of:
- Causing financial or economic damages exceeding $1 billion, or
- Causing the death or severe physical injury of 50 or more individuals.
The bill further narrows regulatory intervention by restricting enforcement to three specific risk categories:
- Biological Weapons: Assisting in the synthesis, engineering, or dissemination of biological agents.
- Loss of Control: A model evading human control or circumventing developer containment safeguards.
- Cyber Warfare: Executing or facilitating severe cyberattacks against critical infrastructure.
Risks falling outside these defined categories, or incidents below the financial and casualty thresholds, would not qualify for federal deployment halts under the proposed statutory framework.
Commerce Department Oversight and Accredited Evaluators
The legislative proposal assigns regulatory execution exclusively to the U.S. Department of Commerce, preventing multiple overlapping federal agencies from independently seeking model injunctions.
Key operational provisions include:
- Centralized Enforcement: The Department of Commerce is designated as the sole federal agency authorized to petition courts to block or pause model deployments.
- Registered Independent Evaluators: The bill establishes a framework for independent testing organizations to register with Commerce and conduct recurring capability audits and safety assessments on frontier laboratories.
- Developer Compliance Obligations: Frontier AI labs would be required to maintain documented safety protocols and submit to accredited third-party evaluations before and during deployment.
Preemption of State Regulations
A central component of Anthropic's proposal is explicit federal preemption over state-level AI safety laws.
Over the past year, state legislatures have introduced divergent compliance frameworks, most notably California's SB 53. By establishing a binding federal standard, the draft legislation would invalidate conflicting state measures in favor of a single national compliance ceiling for frontier developers.
The draft reflects principles Anthropic has advocated since June 2026, translating voluntary lab commitments into federal law while shielding developers from fragmented state-by-state compliance mandates.
Legislative Context on Capitol Hill
Anthropic's draft enters a crowded legislative landscape in Washington:
- The FRONTIER Act: In the House of Representatives, Reps. Lori Trahan (D-Mass.) and Jay Obernolte (R-Calif.) are sponsoring the FRONTIER Act, which similarly explores federal preemption and evaluator standards.
- Bipartisan Senate Negotiations: Senate Majority Leader John Thune alongside Sens. Amy Klobuchar (D-Minn.) and Ted Cruz (R-Texas) continue working on bipartisan legislation targeting catastrophic AI risks.
- Voluntary Accords: The draft follows the multi-lab White House Accord on Super Intelligence signed earlier this autumn by major AI executives, which established high-level safety commitments but lacked binding enforcement mechanisms.
