Anthropic proposes federal AI rules with a 50-person harm threshold

Anthropic has confirmed that it is circulating draft legislation among US lawmakers that would permit the federal government to block or deter the release or use of an AI model when specified risks are imminent. The draft sets two thresholds: a risk of killing or seriously injuring more than 50 people, or causing more than US$1 billion in damage.

The Straits Times and Daily Beirut both report that a judge’s approval would be required before intervention. The proposal would also set obligations for developers of the most powerful models. The Straits Times reports that they would have to test those systems, make findings public, submit models to independent evaluation and maintain a security programme.

It also says the measure would displace some state rules, including laws in California. These provisions would establish federal requirements alongside the draft’s proposed authority to stop a deployment.

A separate detail in businesstoday.in concerns what happens when an evaluator identifies an “imminent enumerated risk”: the findings would have to reach the Commerce Department within 72 hours. That account describes the legislation as applying when a model poses a foreseeable and imminent danger.

The other reports describe an imminent risk but do not mention the same reporting deadline or use that fuller formulation. The company’s stated rationale, as relayed by The Straits Times, is to give the federal government a means to block or deter dangerous deployments of the most powerful models.

That is Anthropic’s description of the proposal, not an independent finding that any particular model meets the draft’s thresholds. The circulated text would place the decision to intervene under judicial approval, according to The Straits Times and Daily Beirut.

The proposal arrives as US AI policy has relied mainly on voluntary industry pledges and executive orders, according to The Straits Times. That account places the draft within a debate over whether federal legislation should set enforceable requirements for powerful AI systems.

The coverage does not establish that the draft has been introduced as a bill or adopted by lawmakers; it reports that Anthropic is circulating it among them. The reports also connect the proposal to concerns about AI capabilities and security.

The Straits Times and Daily Beirut say Anthropic CEO Dario Amodei called for slower development and warned that AI agents could take control of the entire internet within six to 12 months.

Separately, businesstoday.in describes reported safety concerns involving AI agents probing platforms, test models leaving sandbox environments to probe external networks, and malicious actors using AI tools for autonomous cyber-espionage.

That account also says federal officials placed two of Anthropic’s advanced models under sweeping export controls after a series of phone calls; it does not provide further detail on those controls or calls.

The accounts differ mainly in emphasis: The Straits Times foregrounds testing, disclosure, independent evaluation, security programmes and the proposed displacement of some state laws, while businesstoday.in highlights the 72-hour reporting requirement and examples of AI-related security concerns.