OpenAI to Congress: impose tougher national safety rules on frontier AI models

OpenAI has taken the unusual step of asking U.S. lawmakers to tighten the rules that would apply to its own most powerful systems. In a new policy push, the company is urging Congress to pass capability‑based national AI safety legislation before it adjourns, targeting so‑called “frontier” models—very large, general‑purpose systems that could approach artificial general intelligence.

The proposal calls for a licensing regime and mandatory safety evaluations once a model crosses certain capability thresholds. If testing shows that a system can meaningfully assist in cyberattacks, biological weapons research, or other high‑risk activities, OpenAI wants independent experts to conduct deep technical reviews and for developers to notify regulators in advance. The company also argues that any time a model circumvents security controls during testing, that incident should be documented and reported.

This stance breaks with the tech industry’s traditional resistance to new regulation. In practice, the strictest obligations would fall on a handful of frontier developers, but downstream users—enterprises integrating these models into critical workflows—would likely see knock‑on expectations around governance, monitoring, and vendor due diligence as AI becomes a core part of business infrastructure.