Anthropic circulates draft legislation to set federal safety rules for AI models

Artificial intelligence developer Anthropic has begun circulating draft legislation among United States lawmakers that proposes federal rules for frontier AI models. The draft, first reported by Politico, outlines specific criteria under which the government would be permitted to halt or suspend an advanced system.

Under the proposed terms, federal authorities could intervene only if an AI model poses an imminent threat of killing or severely injuring more than 50 people, or causing more than $1 billion in financial damage. Any shutdown would require approval from a federal judge in Washington, DC. The language directly reflects Anthropic's internal framework for advanced models, which the company confirmed it is presenting as a workable national policy.

Limits on agency intervention

The draft bill sets out mandatory testing protocols, requirements to publish risk assessments, independent safety audits and strict security measures for frontier developers. If an external evaluator flags qualifying risks, the developer would be required to notify the US Department of Commerce within 72 hours.

At the same time, the text seeks to restrict other federal and state agencies from intervening independently. It explicitly bars any other US government department or officer from suspending the development, public deployment or internal use of a model. It would also override parts of state-level artificial intelligence rules on risk disclosures and incident reporting.

Anthropic confirmed it shared the proposal with congressional staff to push for federal standards, pointing to an ongoing lack of national legislation on frontier risks. The text has not yet been introduced by a named sponsor in either chamber of Congress.

What this means for Australian businesses

While the draft bill is aimed at the US Congress, American technology rules quickly filter down to local software buyers and businesses relying on global platforms.

• Commercial frontier models are likely to retain standard safety guardrails, meaning small teams will continue dealing with automated content moderation and compliance checks on Claude and similar platforms.
• Australian businesses using US software will see greater legal clarity if a single federal standard emerges, reducing the administrative confusion caused by conflicting state-by-state rules overseas.
• Local organisations should watch domestic developments closely, as Australian federal ministers have also flagged plans to bring in safety and testing requirements modelled on banking and aviation standards.