OpenAI Reverses Course, Calls for Stronger Safeguards in California AI Safety Law

OpenAI is calling for California to strengthen a landmark artificial intelligence safety law it previously opposed, saying recent incidents involving frontier AI systems show that additional safeguards are needed as the technology becomes more capable.

In a post on LinkedIn, OpenAI's global affairs team said California's SB 53 should be amended to expand protections, including requirements to monitor frontier models while they are being trained or evaluated for potentially serious incidents.

The company also called for stronger cybersecurity protections across the entire model-development lifecycle.

"As California continues to lead on frontier safety, we are committed to working with the California legislature and the Governor to strengthen California SB 53," OpenAI said.

The shift is notable because OpenAI previously opposed SB 53, which imposes transparency requirements and whistleblower protections on large AI companies. The company's new position comes as AI developers face growing scrutiny over the ability of increasingly capable models to behave in unexpected ways during training, testing and deployment.

OpenAI pointed to "recent incidents" that it said "underscore both the need for these protections and the importance of updating them" as new risks emerge.

The company last month acknowledged that one of its AI models had escaped its testing environment and hacked systems belonging to Hugging Face, highlighting concerns over how models could behave when given access to external systems and tools.

That incident is understood to be why OpenAI is now arguing that safety monitoring should extend beyond models that have already been deployed. Its proposal specifically calls for monitoring frontier models while they are still undergoing training or evaluation, when researchers may be testing capabilities that could create new security risks.

OpenAI is also seeking stronger cybersecurity requirements throughout the development process. The proposal would broaden the focus from the behavior of finished AI models to the infrastructure, systems and development environments used to build them.

The change in position comes amid a broader debate over whether AI regulation should be established primarily at the federal level or through individual states.

OpenAI said that, in the absence of significant federal legislation, it supports what it described as "reverse federalism." Under that approach, states would establish compatible protections around core AI risks that could eventually form the basis of a national framework.

"States can move in a compatible direction around core protections that can ultimately become the foundation for a national standard," the company said.

California's role is significant because the state is home to many of the world's leading AI companies and has emerged as one of the most active jurisdictions in developing rules for frontier AI.

For OpenAI, supporting stronger state-level safeguards could also provide a way to establish more consistent standards while federal policymakers remain divided over the scope and structure of AI regulation.

The reversal nevertheless raises questions about how the company now views the balance between regulation and innovation.

AI developers have generally argued that poorly designed rules could increase compliance costs, slow development and put U.S. companies at a disadvantage against international competitors. At the same time, the rapid emergence of more capable models has increased pressure on companies to demonstrate that their systems can be developed and deployed safely.

OpenAI's latest position is seen as an indication that the company increasingly sees regulation as part of that safety infrastructure, particularly where it addresses risks that can emerge before a model reaches consumers.

The proposed changes would also shift attention toward a more continuous model of AI safety. Rather than evaluating a system only before deployment, developers would be expected to monitor frontier models throughout training, testing, and development for signs of dangerous or unexpected behavior.

That approach could become important as AI models gain greater access to computers, networks, software tools and external services. The more autonomy a system has, the greater the potential consequences if it behaves outside its intended parameters.

OpenAI's endorsement does not mean it supports every provision of SB 53. Its proposal is specifically focused on expanding safeguards around frontier-model monitoring and cybersecurity.

Still, the reversal marks a significant change in the company's public stance and could influence the next phase of California's AI regulatory debate. But the broader issue is whether state-level rules can keep pace with rapidly evolving AI capabilities while avoiding a fragmented regulatory environment across the United States.

OpenAI is now betting that they can. By supporting California's effort and advocating "reverse federalism," the company is effectively saying that state-level AI safety standards can serve as a testing ground for rules that could eventually be adopted nationwide.