OpenAI pushes frontier AI rules from state experiments to federal law

Frontier AI governance is moving from voluntary lab policy toward enforceable institutions. In a June 3, 2026 blueprint, OpenAI argues that the United States should turn the emerging state-level consensus on frontier AI safety into a national framework.

The document starts with California’s SB 53, New York’s RAISE Act, and Illinois’s SB 315. OpenAI says those state efforts have converged around a practical baseline: severe-risk evaluations, transparency reports, independent auditing, critical safety incident reporting, model-weight security, whistleblower protections, and enforceable accountability. Its notable framing is “reverse federalism”: states develop common governance patterns first, Congress then converts that work into a federal framework, and overlapping state rules for the same frontier safety risks are preempted after the national standard exists.

The second pillar is CAISI, the Center for AI Standards and Innovation. OpenAI wants CAISI authorized and funded as the federal government’s main frontier AI evaluation institution, with access to cyber, CBRN, national-security expertise, and classified compute environments. The strongest operational proposal is a mandatory CAISI evaluation for the most capable frontier models before public release. OpenAI stops short of making CAISI a deployment gatekeeper: the agency would evaluate and recommend mitigations, while developers would remain responsible for deployment decisions and public disclosures.

The third pillar is resilience across government. OpenAI argues that frontier AI should be treated as a national priority spanning public health, cybersecurity, scientific agencies, diplomacy, economic policy, and international safety coordination. The blueprint calls for legal certainty around safety collaboration, tighter protection of America’s compute advantage, limits on unevaluated frontier systems in sensitive government contexts, and investments in AI-enabled biodefense, cybersecurity, critical infrastructure protection, and rapid response.

This is not law, and OpenAI is an interested party in the debate. Still, the policy signal matters. A leading model developer is publicly endorsing independent audits, incident reporting, model-weight security duties, and government-led evaluations in one package. The next test is whether Congress gives CAISI real authority and budget, and whether other frontier labs accept a similar level of external scrutiny.