Anthropic pitches AI safety bill that would limit federal model shutdowns

Anthropic has privately pitched lawmakers an AI safety bill that would impose new oversight duties while sharply restricting federal power to suspend advanced models. POLITICO disclosed the effort on October 16, 2026. The confidential draft would reserve suspensions for specified imminent catastrophic threats and require a federal judge’s approval—not simply an agency’s decision.

POLITICO’s reporting cited three people familiar with the discussions, according to Law Commentary. An Anthropic spokesperson confirmed that the draft reflects proposals in the company’s Advanced AI Framework. The proposal has not been formally introduced in Congress, so its provisions describe the rules Anthropic is seeking, not powers or obligations already in force.

A narrow route to suspension

The draft targets frontier models: highly advanced AI systems capable of increasingly complex tasks. Federal officials could seek to suspend a covered model only for a foreseeable and imminent threat involving more than $1 billion in damages, or the deaths or serious injuries of more than 50 people. Those are alternative thresholds; the proposal would not require both.

Scale alone would not qualify a threat for intervention. It would also have to fall into one of three categories:

• The creation or release of biological weapons.
• Cyberattacks.
• AI systems escaping their developers’ control.

Other forms of harm would sit outside this specific suspension framework. That distinction separates the draft’s general safety requirements from its emergency intervention power: an obligation to assess or disclose risk would not, by itself, make every identified risk grounds for stopping a model.

Even a qualifying threat would not let the Commerce Department independently shut down a model. It would need an order from the U.S. District Court for the District of Columbia. The draft would also bar other federal departments, agencies and officials from independently suspending a covered model’s development, deployment or internal use through the proposed framework.

Outside scrutiny, with company-paid evaluators

On the oversight side, the draft would require outside evaluators to register with Commerce and periodically assess developers and their frontier models. Developers could pay the organizations inspecting them, provided the compensation did not depend on their findings. The proposed independence rule therefore concerns how evaluators are paid, rather than prohibiting developers from funding the reviews.

The duties would extend beyond outside assessments. Developers would have to establish safety frameworks, publish transparency reports for new models, protect whistleblowers and report critical incidents to Commerce, according to Interesting Engineering’s account of the draft. These requirements cover both a developer’s safety practices and information about its models and incidents.

Federal duties could replace some state rules

The proposed national requirements would not simply sit alongside every state rule. The draft would override some state requirements covering public risk disclosures, independent evaluations and critical-incident reporting. That mechanism, known as federal preemption, would let the federal framework displace state obligations in the areas it covers.

The extent of that displacement would depend on the final legislative language. For developers, the proposal offers a federal set of duties paired with limits on who can order a suspension. For states, it could restrict separate safety requirements in those same areas. The draft thus addresses two different boundaries: when Washington could stop a model, and which safety obligations states could continue to impose.