Artificial intelligence company Anthropic is facing criticism over a proposed AI safety bill that would give the U.S. government authority to intervene when advanced AI models pose serious threats involving biological weapons, cyberattacks, or widespread financial damage.
According to an October 16 report by Politico, Anthropic presented lawmakers with a confidential proposal outlining specific thresholds for government intervention. Under the reported framework, federal authorities could seek to halt the release or use of a frontier AI model if it presents a foreseeable and imminent risk of causing more than 50 deaths or serious injuries, or exceeding $1 billion in damages, through certain AI-enabled activities.
The proposal focuses on threats involving AI-assisted biological weapons, models circumventing developer security controls, and sophisticated cyberattacks. However, critics argue that the proposed thresholds could leave significant AI-related threats outside the bill’s emergency intervention provisions. A biological attack resulting in fewer than 50 deaths or serious injuries, for example, might not independently meet the reported threshold for intervention under this particular framework.
The legislation would also introduce independent evaluations of frontier AI developers and their models. Evaluators, potentially compensated by the AI companies themselves, would periodically assess whether advanced systems pose qualifying security risks. If an imminent threat is identified, the U.S. Department of Commerce could seek a court order preventing the model’s release or use. The reported proposal would assign this specific intervention authority to the Commerce Department rather than other federal agencies.
Despite Anthropic’s stated commitment to responsible AI development, the proposal has generated controversy among AI researchers and policy advocates. Heidi Klaaf, Chief AI Scientist at the AI Now Institute, criticized the approach as autocratic, arguing that allowing an AI company to propose thresholds for acceptable risk could undermine public accountability and overlook existing harms associated with artificial intelligence. The reported draft is not enacted law, and questions remain about whether its proposed thresholds provide sufficient protection against emerging AI cybersecurity threats.