Anthropic urges binding tests for frontier AI models

Anthropic is urging policymakers to move from transparency measures to binding oversight for frontier AI models, arguing that recent cybersecurity capabilities in Claude Mythos Preview show that advanced systems have become strategic tools with public safety implications. The proposed framework would require third-party testing for high-compute models in cybersecurity, biological weapons, loss of control, and automated R&D, with government power to block or reverse deployment when risks are unacceptable.

The policy agenda also calls for governments to prepare for AI-driven labor disruption through better measurement, pro-employment incentives, and possible long-term income support if demand for labor falls. At the same time, regulators in fields such as biomedicine are urged to adapt approval systems so AI-accelerated drug discovery, simulations, and clinical methods do not get delayed by processes built for slower innovation.

The broader plan frames AI as a challenge for civil liberties and geopolitics. It recommends safeguards around autonomous weapons, surveillance data, and access to AI during government action, while calling for a democratic coalition to coordinate safety rules, supply chains, export controls, benefit-sharing, mutual defense, and resistance to AI-enabled repression.