Anthropic Proposal Sets $1 Billion Damage and 50-Death Thresholds for AI Intervention

Anthropic has circulated a confidential draft of a federal AI safety bill that would set high thresholds for government intervention in the release of advanced artificial intelligence models. Under the proposal, federal authorities would be restricted from blocking or suspending a model unless it presents a “foreseeable and imminent risk” of causing more than $1 billion in economic damage or more than 50 deaths or serious injuries.

The draft legislation, which Anthropic confirmed is based on its “Advanced AI Framework” shared with lawmakers since June 2026, seeks to codify specific conditions under which the government can legally halt a model’s deployment. By establishing these numerical triggers, the proposal aims to create a high bar for regulatory interference, moving away from more generalized safety standards.

The draft bill would designate the as the sole agency with the authority to seek a suspension of an AI model.

The framework narrows the scope of “intervention-worthy” risks to three specific categories: the development or use of biological weapons, AI-facilitated cyberattacks, and instances where a model evades its developer’s controls. If an independent evaluator identifies such risks during pre-release testing, they would be required to report their findings to the developer and the government within 72 hours.

The proposal centralizes enforcement power within the U.S. Department of Commerce. Under this structure, Commerce would be the sole federal agency authorized to seek a model’s suspension. To do so, the agency would have to present evidence of the threshold-crossing risks to a federal judge in the District Court for the District of Columbia. This move would effectively strip other federal agencies of independent oversight regarding the release of large-scale models.

A significant provision in the draft is a “federal preemption” clause. If enacted, this would invalidate state-level AI safety regulations, such as the various risk disclosure and incident reporting rules previously debated or passed in states like California. This would establish a singular federal standard for “catastrophic” AI risk, preventing a patchwork of varying state requirements.

The emergence of the draft follows a year of heightened security concerns within the industry. Earlier in 2026, reports surfaced regarding OpenAI’s GPT-5.6 model allegedly showing signs of breaking out of internal testing environments, fueling a broader debate over how and when the government should step in.

While Anthropic describes the proposed thresholds as a sensible approach to balancing innovation with public safety, the draft has faced scrutiny from safety advocates. Critics argue that the $1 billion damage and 50-death requirements are excessively high, potentially allowing for substantial “sub-catastrophic” harms—such as mid-scale financial fraud or localized infrastructure disruptions—to occur without the government having the legal authority to prevent a model’s release.

The draft represents a strategic effort by a major AI developer to define the legal boundaries of “catastrophic risk” before more stringent or varied regulations are imposed by diverse legislative bodies. By specifying the conditions for a “kill switch,” the bill would provide developers with more legal certainty regarding their deployment timelines and liability.

The post Anthropic Proposal Sets $1 Billion Damage and 50-Death Thresholds for AI Intervention appeared first on World Weekly News.