Anthropic on Wednesday set out two policy proposals aimed at regulating advanced artificial intelligence as capabilities accelerate. The company's centerpiece recommendation, titled the Advanced AI Framework, outlines legal mechanisms for governments to prevent or punish deployments deemed dangerous.
Under the framework, regulators would be empowered to impose civil penalties tied to a developer's global annual revenue. Those penalties would be structured to rise with repeat infractions, creating a financial deterrent to unsafe deployments.
The proposal targets a specific threshold of technical scale and market presence. It applies to AI systems trained with more than 10²⁵ floating-point operations and to organizations that either generate over $500 million in AI-related revenue or spend more than $1 billion on AI research and development.
Anthropic identifies four categories of catastrophic risk that the framework aims to address:
- Biological threats arising from AI-assisted development of weapons or other harmful agents;
- Cyber vulnerabilities at scale that could be exploited across critical systems;
- Loss of control over AI systems; and
- Automated research and development activities that could magnify other hazards.
To illustrate the potential scope of risk, the company referenced its Claude Mythos Preview model, noting that the system this year identified thousands of high-severity vulnerabilities across major operating systems and browsers.
The Advanced AI Framework would require developers of frontier models to meet several compliance measures. Those include systematic testing of models with publication of results, transparency about safety practices, submission to independent evaluation, and maintenance of strong security programs to mitigate cyber threats.
Alongside technical and regulatory proposals, Anthropic published an Economic Policy Framework that addresses workforce preparation and the distribution of financial benefits stemming from AI progress. The company framed this as a complement to the safety-focused measures.
Anthropic directed its recommendations principally at the U.S. federal government. At the same time, the company stated that Congress should not preempt state laws unless any federal legislation is at least as robust as the suggested framework. The proposal also allows states to continue regulating matters such as child safety and consumer protection where those issues fall outside the specific safety functions covered by federal law.
The policy documents propose a mix of preventative controls, disclosure obligations and independent oversight aimed at containing high-consequence risks as AI systems scale in capability and deployment.