Anthropic has privately submitted a draft federal legislative framework to members of the US Congress to govern the safety and security of frontier artificial intelligence models, according to a report by Politico. Under the proposed bill, the federal government could block the release or use of an advanced model only under strict numerical and judicial conditions. Specifically, federal intervention would require a clear and imminent threat involving biological or chemical weapons, a total loss of developer control, or cyberattacks projected to cause at least 50 deaths or severe injuries, or more than $1 billion in property damage. Enforcement would be centralized solely within the US Department of Commerce, stripping other federal agencies and state authorities of independent shutdown powers, while requiring an order from a federal district judge in Washington before any deployment halt could take effect.
The initiative stands in sharp contrast to precautionary frameworks such as California's SB 1047 legislative effort and the European Union AI Act, which granted broader executive discretion to intervene before catastrophic events unfold. Anthropic's proposal establishes mandatory periodic audits by independent evaluation entities registered with the Commerce Department, requiring critical hazards to be reported within 72 hours. However, the measure also permits AI developers to directly fund these external testing bodies, drawing criticism over financial conflicts of interest. An Anthropic spokesperson described the text as a practical federal version of the advanced safety framework the company has circulated to federal and state lawmakers since June.
Civil society advocates and policy groups swiftly condemned the draft, arguing that its steep casualty threshold creates a legal shield for tech companies while leaving low-intensity harms unaddressed. Sneha Revanur, head of Encode Action, warned that the government must preserve emergency authority without being forced to prove whether an incident will kill 50 people or 49, characterizing the effort as self-serving corporate lobbying. Brad Carson, president of Americans for Responsible Innovation, stated that the bill falls short of minimum federal safety nets by allowing developers to set and dilute their own safety standards. Over the next six months, lawmakers and competing frontier model developers are expected to clash over whether to adopt these judicial pre-authorization rules or retain independent agency powers.