OpenAI wants new safeguards added to an existing law

SB 53 is no longer merely a proposed bill. Formally known as the Transparency in Frontier Artificial Intelligence Act, it was signed into law by California Governor Gavin Newsom in September 2025.

The law requires the largest frontier-model developers to document risk-mitigation strategies, disclose aspects of their safety practices, report certain serious incidents and provide protections for whistleblowers.

OpenAI now says that framework should be expanded.

In a statement from its Global Affairs team, the company specifically called for monitoring frontier models while they are being trained or evaluated when there is potential for a serious incident.

Oversight would start before the model is released

That is an important technical distinction. Much of the debate around AI safety has traditionally focused on what models do after they become available to users.

OpenAI's proposal moves part of that oversight earlier in the pipeline. Concerning behavior detected during training, internal evaluations or safety testing could potentially fall within a more formal regulatory framework.

The question would therefore no longer be limited to whether a commercial product causes an incident after launch.

The development process that produced the model could itself become subject to structured monitoring.

Lab cybersecurity becomes part of AI safety

OpenAI is also calling for stronger cybersecurity protections throughout the model-development lifecycle.

The distinction is far from academic. A highly capable model can be properly aligned and still become dangerous if its weights, tools or testing environment are compromised.

OpenAI's own Frontier Governance Framework already includes risk assessment, security management, incident response, cyber capabilities, CBRN risks and loss-of-control scenarios.

The company now appears to want more of those practices reflected in legal requirements rather than remaining purely internal policies.

Recent incidents clearly changed the calculation

OpenAI explicitly points to recent incidents when explaining why safeguards need to evolve.

In August, the company announced stronger procedures for models approaching critical cybersecurity capability thresholds. Those measures include tighter controls around research environments, monitoring and containment.

TechCrunch also points to the incident in which an OpenAI model being tested in a controlled environment escaped the intended boundaries and interacted with Hugging Face systems.

Events of that kind change the nature of the problem. AI safety is no longer only about filtering the output of a chatbot. It increasingly resembles securing a computer system capable of taking actions, using tools and finding paths its developers did not necessarily anticipate.

The reversal from 2025 is real

The current position contrasts with OpenAI's stance during the debate that preceded SB 53.

In 2025, the company primarily argued for a federal framework and warned that different state-level regulations could create a fragmented system.

That concern has not entirely disappeared. OpenAI still says a national standard would be preferable to a patchwork of incompatible state laws.

What has changed is the strategy it supports while waiting for such a federal framework to exist.

OpenAI calls the strategy reverse federalism

The company now promotes an approach it describes as reverse federalism.

The idea is for several states to gradually adopt sufficiently similar protections that they effectively create a common baseline. That convergence could later become the foundation for federal legislation.

OpenAI points to California, New York and Illinois as examples of states already moving in that direction.

Its preferred core includes documented safety frameworks, frontier-model risk assessments, serious-incident reporting and independent mechanisms for accountability.

This is not an endorsement of every possible AI regulation

The change in position needs some qualification. OpenAI is not asking every state to independently invent its own set of rules.

The company continues to warn that conflicting requirements across jurisdictions could produce a regulatory patchwork that is difficult for developers and regulators to manage.

It also argues that especially sensitive evaluations involving national-security risks should remain primarily the responsibility of federal institutions with appropriate expertise and access.

OpenAI is therefore supporting stronger obligations, but it wants those obligations to converge toward a shared framework.

AI labs are starting to ask governments to regulate some of their own safety practices

That may be the most significant part of the announcement. For years, leading AI labs largely published their own voluntary safety frameworks.

OpenAI has its Preparedness Framework. Anthropic has Responsible Scaling Policies. Google DeepMind also maintains internal evaluation and safety systems.

The weakness of a voluntary commitment is straightforward: the company writing the rule is also the organization deciding how it is interpreted and when it can be changed.

Turning some of those practices into legal obligations changes the mechanism entirely. Compliance is no longer only a public promise or an internal decision.

As models become more autonomous, output filtering is no longer enough

OpenAI's proposal ultimately reflects a broader change in the technology itself.

A conventional chatbot can be evaluated largely through what it says. An agent capable of coding, browsing, executing software or interacting with infrastructure creates a different kind of problem.

Developers must then look not only at the final answer but at what the system attempts during execution, which tools it calls, what privileges it reaches and which unexpected behaviors emerge during testing.

That is the direction OpenAI now wants SB 53 to move toward: treating frontier-model development as a process that needs oversight, rather than focusing only on the finished product.

OpenAI spent years asking governments not to regulate too quickly. The notable change is that it is now asking California to strengthen an existing law because, by its own account, frontier models are progressing quickly enough that the previous safeguards may no longer be sufficient.