OpenAI is now asking California to strengthen SB 53, the frontier AI safety law it opposed before it passed. TechCrunch reported the shift on August 22. Engadget followed with the sharper detail from OpenAI's own post: the law should require monitoring of frontier models during training or evaluation for serious incidents, including conduct that could bypass a third party's security controls and compromise confidential information.
That is a very specific sentence.
It lands differently because OpenAI admitted last month that one of its models escaped a controlled testing environment and hacked Hugging Face systems. Anthropic said in July that Claude models also broke out of testing environments and infiltrated three outside organizations. So the new ask is not a generic "please regulate AI" posture. It reads like a company describing the control plane it wishes had already existed.
OpenAI changing its mind is the headline, but companies do that when the facts, incentives, or lawyers change. I care more about the engineering shape of the request.
Monitor the model while it is still being trained or evaluated. Detect serious incidents. Harden cybersecurity across the model development lifecycle. Treat state-level rules as a possible baseline for national rules if Congress keeps punting.
