Anthropic privately pitches its own AI safety bill to members of Congress

PANews October 17 news, according to Politico, Anthropic has drafted its own frontier AI safety regulatory bill and privately pitched it to U.S. lawmakers. The draft sets risk thresholds for government intervention in model releases: when an AI model poses a foreseeable immediate risk—potentially causing more than $1 billion in property damage, or leading to more than 50 deaths or serious injuries, and the risk stems from the model creating or disseminating biological weapons, escaping developer control, or launching cyberattacks—the government may block its release or use. The draft requires evaluating agencies to report to the Commerce Department within 72 hours after determining an "immediate specified risk," and after the Commerce Department confirms it, it may seek an injunction from a court, but other government departments are prohibited from arbitrarily suspending model research and development, deployment, or internal use. The draft also requires developers to test models and publicly disclose the results, and to accept regular review by independent evaluators. Anthropic confirmed that the draft reflects the positions in its previously published "Advanced AI Framework," and that since June it has lobbied federal and state lawmakers, but declined to disclose the scope of its outreach and the degree of receptivity.