Anthropic has unveiled a sweeping policy proposal urging governments to gain legal authority to block or deter high-risk artificial intelligence launches. The company’s "Policy on the AI Exponential" plan, outlined by CEO Dario Amodei, argues that AI is advancing far faster than regulatory systems can keep pace and that transparency alone is no longer sufficient for frontier models.
The Advanced AI Framework targets models trained above 10²⁵ floating-point operations and companies with more than $500 million in AI-related revenue or $1 billion in AI R&D. It identifies four critical risk areas: biological threats, cyber risks, loss of control, and automated AI research. Under the plan, civil penalties would be tied to global annual revenue, with higher fines for repeat violations. Developers would be required to test models before release, publish safety summaries and system cards, and undergo independent evaluation. Security programs would need to protect model weights and training environments from both external and insider threats.
Alongside the safety framework, Anthropic released an Economic Policy Framework aimed at preparing workers for AI-driven job displacement and sharing financial benefits. Amodei stressed that both economic provision and human purpose must be addressed, calling job displacement "undesirable and dangerous."
The proposal comes as Anthropic expands access to its restricted Claude Mythos model, with variants capable of executing complex autonomous cyber attacks. Amodei compared the needed oversight to the FAA’s role in aviation, stating that "frontier AI models, like airplanes, should be required to go through technical testing and auditing, and their release should be blocked or reversed as a threat to public safety." The essay acknowledges the Trump administration’s executive order but pushes for stronger action.
OpenAI CEO Sam Altman previously accused Anthropic of using "fear-based marketing" to promote its models and justify concentrated control. Amodei pushed back, saying public concern reflects legitimate risks and that transparency is a democratic accountability mechanism. He urged policymakers to act before capabilities outstrip all guardrails, calling for joint government-industry work on detecting and containing unsafe systems.
