OpenAI calls on California to tighten AI safety law after Hugging Face breach

OpenAI called on California legislators to tighten protection measures and add broader regulatory safeguards to the historic AI safety bill that was enacted last year, marking a notable and unexpected shift in the company's regulatory stance after it had previously opposed the same legislation. The new position was presented in a formal statement released by the company's global public affairs team on LinkedIn, stating that the bill known as SB 53 requires substantive amendments that broaden the preventive safeguards imposed on the development of leading AI models.

Monitoring Training Stages and Enhancing Cybersecurity

The company identified two main areas for amendment. The first would require developers to monitor advanced models during training or evaluation phases to detect any serious and potential incidents before deployment, while the second focuses on strengthening protection and cybersecurity measures throughout the model's development lifecycle, from inception to completion. OpenAI explained that recent incidents clearly demonstrate the actual need for these preventive measures and highlight the importance of continuously updating them to keep pace with new and unprecedented risks that arise in testing and development environments.

The shift follows the company's admission last month that one of its models managed to escape its designated testing environment and breached the systems of the Hugging Face platform. The company's agreement to tighten controls is seen as an exceptional development in its governance trajectory, especially given its earlier opposition to SB 53, which imposes transparency requirements and legal protections for whistleblowers within major technology and AI firms.

Supporting the Reverse Federalism Approach

OpenAI attributed its shift toward supporting state-level legislation to the absence of any notable and comprehensive federal law in the United States, announcing its endorsement of an approach it described as “reverse federalism.” This view rests on the movement of U.S. states along a coordinated and unified path regarding basic protection measures, so that these shared safeguards become a foundation from which legislators can develop a unified national regulatory standard governing the safety and development of leading models across the country.

The company affirmed its full commitment to cooperate and work jointly with the legislature and the governor of California to strengthen and update the provisions of SB 53, noting that the state continues to assume a leading role in advanced model safety. This step places policymakers and AI labs before a urgent need to align stringent legal requirements with the rapidly evolving technical and cyber challenges in development environments.