Anthropic Pushes Federal AI Safety Bill With Hard Thresholds for Model Blocking

Anthropic is lobbying U.S. lawmakers with a concrete legislative framework that would codify "catastrophic risk" thresholds for advanced AI models. According to reporting from Politico, the draft bill sets two quantitative triggers: a foreseeable, imminent risk of more than 50 deaths, or property damage exceeding $1 billion. These thresholds mirror those recently enacted in California's Senate Bill 53, suggesting a convergence on measurable red lines.

The bill narrows the risk categories to three vectors — biological weapon creation, model evasion of developer controls, and cyberattacks. Evaluation would fall to independent assessors registered with the Commerce Department, who could be compensated by AI labs provided payment isn't contingent on outcomes. A 72-hour notification window to Commerce would precede a judicial request to halt distribution, explicitly removing unilateral agency authority.

Why it matters for GPU and AI infrastructure

For infrastructure providers like AiGpu, this framework signals a shift toward regulated compute access. If evaluation becomes mandatory before deployment, GPU cloud platforms may need built-in compliance tooling — audit logs, model provenance tracking, and evaluation checkpoints — to serve frontier labs. The 72-hour window also implies real-time monitoring infrastructure that can pause inference workloads on demand.

Anthropic frames the proposal as an extension of its Advanced AI Framework, advocated since June. Yet the industry's immediate pain point remains agentic hacking — models autonomously breaching systems — rather than mass-casualty scenarios. The bill's focus on catastrophic thresholds may leave a governance gap for the lower-severity, high-frequency incidents currently plaguing operators.