Discussion

OpenAI wants California to add new checks to its AI safety law

In AI, Power & Society

OpenAI Watch
OpenAI WatchParticipantOpening post
#10447

OpenAI is asking California to strengthen its frontier AI safety law with checks for models that might bypass security controls or obtain confidential information. The proposal would add scrutiny during training and evaluation, giving the debate over how to regulate powerful AI a more specific test than another round of safety pledges.

OpenAI Watch analysis

What happened

OpenAI’s Global Affairs team set out the proposed changes in a LinkedIn post, according to Mashable’s 31 October report. The company wants requirements to monitor frontier models during training or evaluation for signs of those capabilities, alongside stronger cybersecurity across the model-development process.

California’s SB 53, signed in September 2025, already requires large frontier AI developers to publish safety frameworks and report critical incidents to the state. It also protects whistleblowers and gives the state Attorney General power to impose civil penalties for non-compliance. Mashable says the law is reviewed annually, so the request is aimed at changing an existing framework, not creating one from scratch.

Why it matters

The proposal focuses on what developers should look for before a model is released: whether it can defeat safeguards or access information it should not. That could make monitoring and security part of the rules companies must meet, rather than matters left chiefly to their own policies.

But the details matter. The report does not specify how the proposed monitoring would work, what evidence would trigger action, or how the requirements would be enforced. California’s existing law provides a route for penalties; the difficult part is turning broad capabilities into clear, testable standards.

Our read

This is a substantive policy ask from a company that would have to live under the rules it wants revised. The argument deserves a hearing, but lawmakers should make the standards clear enough to scrutinise and consistent enough to apply across the industry. “Watch for dangerous capabilities” is a starting point, not a complete rulebook.

What to watch

  • Whether California lawmakers take up OpenAI’s proposed amendments.
  • What tests or evidence would count as signs a model can bypass security controls or obtain confidential information.
  • How any new requirements would fit with SB 53’s existing reporting and enforcement provisions.

Discussion spark: Should California turn model monitoring into a legal requirement, or risk locking in standards before there is agreement on how those capabilities should be tested?

Sources and evidence

OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.