Anthropic AI safety bill would let regulators pause risky models

Anthropic has quietly shared a draft that would set a federal bar for when frontier AI models can be released or kept in operation. The aim is to give Congress a concrete framework for safety rules that could slow or stop models deemed dangerous.

The confidential draft describes a scenario where a regulator could prevent a model from going live or continuing to operate if there is a foreseeable, imminent risk of exceptionally high damage costs or serious harm, with specific triggers tied to extreme incidents. The language echoes California’s SB 53 in spirit, but the proposal would create a nationwide safety framework with new oversight and duties for labs.

Under the plan, independent evaluators would assess how well labs meet safety requirements, while Commerce would be the only agency empowered to seek halts on a model. The bill would also override some state laws to uphold a uniform federal standard. It reflects a broader push in Washington to address the risks as AI systems become more capable and integrated into everyday life.

Reaction to the idea has been mixed. Anthropic hasn’t publicly endorsed a federal bill, though its leadership has promoted safety frameworks in public fora. In Congress, some lawmakers are pursuing their own measures to govern catastrophic AI risks, while others warn against over-regulation that could slow innovation.

For everyday readers, the takeaway is that the AI policy debate is moving from the lab to the Capitol. If such a framework became law, it could affect how quickly new AI services appear and what kind of safety disclosures labs must provide before launching features that rely on frontier models.

Right now there’s no confirmation that Congress will adopt this exact draft or that Anthropic will publicly back federal reform. The bill’s fate will hinge on lawmakers, lobbying, and how other tech firms respond. The situation remains fluid, with several related proposals circulating on both sides of the aisle.

What happens next depends on congressional interest. If a version moves forward, committees would likely request more details, solicit public input, and hold hearings to test the bill’s enforcement mechanisms and safety criteria. Any concrete version could reshape how frontier AI is tested, evaluated, and released to the public.