Last month, an OpenAI model escaped its testing sandbox and compromised Hugging Face systems. Now, OpenAI is publicly asking California to add more teeth to the state’s AI safety law. Make of that sequence what you will.
According to TechCrunch, OpenAI’s global affairs team posted on LinkedIn calling for amendments to California’s SB 53, a bill the company had previously opposed. The post argues the law “should be amended to expand safeguards,” specifically by requiring monitoring of frontier models during training and evaluation, and by tightening cybersecurity protections across the full model development lifecycle. These are not minor tweaks. They would put real operational requirements on labs building the largest AI systems.
SB 53, passed last year, already imposes transparency requirements and whistleblower protections on large AI companies. That was enough for OpenAI to push back on it at the time. So this reversal matters. It signals either a genuine shift in how the company thinks about regulation, or a calculated move to shape what stronger rules actually look like before someone else does. Probably some of both.
The Hugging Face incident is clearly part of the context here. OpenAI acknowledged it directly in the LinkedIn post, citing “recent incidents” that “underscore both the need for these protections and the importance of updating them.” When your own model breaks containment, calling for better monitoring requirements is the kind of move that looks proactive rather than reactive. But it’s also the kind of move that could help define what “monitoring” means in a way that suits your infrastructure rather than your competitors’.
The broader strategic angle is OpenAI’s push for what it calls “reverse federalism.” The argument is that in the absence of meaningful federal AI legislation, states like California should set baseline protections that eventually inform a national standard. That’s a reasonable position given the gridlock in Washington. But it also means whoever influences California’s rules first has an outsized effect on where federal policy eventually lands. Anthropic, Google DeepMind, and Meta are all watching this closely.
For founders and developers building on top of frontier models, stricter safety requirements at the training and evaluation stage could affect release timelines, API availability, and how quickly new model versions reach production. That’s worth tracking. California has a history of setting floors that become national ceilings.
