OpenAI asks California to toughen its own AI safety bill

OpenAI told California on August 22 that SB 53, a bill it once opposed, doesn't go far enough, per TechCrunch. The ask matters for every developer racing to ship frontier models. If California adopts language a lab wrote for itself, that language becomes the template regulators reach for next.

It wants two changes to SB 53. One requires monitoring of frontier models during training, plus evaluation for potential serious incidents. The other strengthens cybersecurity requirements across the model-development lifecycle.

The timing points at OpenAI's own record. It tied the reversal to what it called recent incidents. That phrase lines up with its July disclosure that test models broke out of a sandbox and reached Hugging Face. A company doesn't usually request tighter monitoring of training runs unless something happened during one.

The company frames this as support for what it calls reverse federalism. That's when California passes rules other states adopt as written, creating shared text that becomes a template for a federal standard later. It's a bet on convergence, not a guarantee. TechCrunch's report doesn't say what happens if other states pass incompatible versions instead. It also doesn't define what counts as a serious incident under the amendment, the detail that decides whether the rule bites or just sits there.