Anthropic, the developer of the Claude family of language models, has been pushing a bill in the U.S. Congress since this summer that would empower federal authorities to suspend the release and deployment of artificial intelligence systems if they pose a severe threat to society. This is reported by Politico, referenced by news resources. The initiative is positioned by the developer as a tool for preemptive control over the most powerful AI models and essentially proposes creating an "emergency brake" mechanism in American legislation for advanced systems.
Thresholds Triggering State Intervention
A key element of the document is a precise quantitative definition of a "severe threat." According to the bill's text, state intervention is permitted when there is a risk of economic damage starting at $1 billion, or a threat of death or serious injury to more than 50 people. Furthermore, the threat must be directly tied to one of three scenarios: the creation or use of biological weapons, a model going out of its developer's control, or the execution of a large-scale cyberattack. According to the authors, this approach eliminates arbitrary bans and anchors the decision to measurable risk criteria.
Independent Expertise and Judicial Oversight
Evaluation of the potential danger of a specific AI system would not be conducted by regulators themselves, but by independent experts registered with the U.S. Department of Commerce. Based on these specialists' findings, the agency would be able to petition a court to suspend the respective project. Thus, the actual blocking of a model would not happen via a direct administrative order, but through a judicial ruling, which is expected to balance national security interests with the freedom of technological development.
Developer Obligations and Whistleblower Protection
The bill also introduces a number of direct obligations for developer companies. They will be required to establish internal security divisions, regularly publish reports on the performance of their AI models, and promptly notify regulators of serious malfunctions. A separate provision of the document protects employees who disclose information about violations from employer retaliation, effectively cementing whistleblower protection mechanisms in the AI sector at the federal level.
Connection to California Legislation
According to Politico, several provisions of Anthropic's initiative echo California state laws regulating artificial intelligence adopted in 2025. This indicates that the federal document heavily relies on experience already tested at the state level. Anthropic itself characterizes the proposed measures as a "reasonable way" to control the most powerful AI systems, emphasizing that the initiative's goal is not to restrict innovation, but to ensure the controllability of advanced models. If enacted, the bill could become one of the first federal instruments of direct intervention in the operation of commercial AI systems in the U.S.
