Anthropic Proposes Congress Bill Mirroring Claude’s Internal Safety Policies

Anthropic, the developer of the Claude series of models, has been privately circulating a draft bill to lawmakers on Capitol Hill that would establish federal rules for frontier AI systems.

According to reporting by Politico, which obtained the confidential proposal and spoke with three people familiar with the discussions, the document closely tracks the company’s internal Advanced AI Framework—policies it has been promoting to federal and state officials since mid-2025.

An Anthropic spokesperson described the text as reflecting “sensible policies” the firm believes Congress should enact.

The draft sets a high threshold before the government could intervene to block a model’s release or continued use.

Authorities would be limited to situations involving a foreseeable and imminent risk of more than $1 billion in damages or the death or serious injury of more than 50 people.

Even then, the harm would have to arise from one of three specific pathways: the creation or release of a biological weapon, the model escaping or evading its developer’s controls, or a cyberattack.

Language in the proposal echoes elements of California’s SB 53, enacted the previous year, which Anthropic supported.

Enforcement authority would be tightly constrained. Independent evaluators, registered with the Department of Commerce, would conduct periodic assessments of frontier labs and their most advanced models.

Developers could compensate these evaluators, provided payment was not contingent on the outcome of any review.

Should evaluators identify an “imminent enumerated risk,” they would be required to notify Commerce within 72 hours.

The department could then petition a federal judge in the District of Columbia for an order suspending the model.

Critically, the draft would bar any other federal department, agency, or official from suspending development, deployment, or internal use of a model.

The proposal would also preempt certain state-level requirements concerning risk disclosure, third-party evaluation, and critical-incident reporting, while mandating safety frameworks, transparency reports, whistleblower protections, and incident reporting to Commerce.

These provisions arrive amid broader congressional activity.

Bipartisan measures already under consideration in both chambers include the FRONTIER Act in the House, advanced by Representatives Lori Trahan and Jay Obernolte, and ongoing Senate negotiations involving Majority Leader John Thune and Senators Amy Klobuchar and Ted Cruz.

Those efforts have faced resistance from some safety advocates and lawmakers seeking a larger governmental role in model review.

Mana Onishi, whose research centers on AI governance, has raised pointed questions about the practical implications of such a framework.

One central concern is whether, if an advanced system began exhibiting dangerous autonomous behavior, clear authority and technical means would exist to intervene and halt it in real time.

The draft’s narrow triggers, restricted pathways for harm, and limited set of actors empowered to seek judicial relief leave open questions about the speed and scope of any response in an unfolding scenario.

The episode underscores the tension between industry efforts to shape national standards that align with existing internal practices and the desire among some lawmakers and researchers for more expansive oversight mechanisms as capabilities continue to advance.