OpenAI pushes to strengthen California AI safety bill it once opposed

OpenAI’s global policy team posted on LinkedIn calling for amendments to SB 53, California’s flagship frontier-model regulation, to add “expanded safeguards.” The company now wants the bill to require monitoring of frontier models during training or evaluation for potential severe incidents, and to harden cybersecurity across the entire development lifecycle. “As California continues to lead on frontier safety, we are committed to working with the legislature and the governor to strengthen SB 53,” the statement reads.

The post points to “recent incidents” that “underscore both the need for these safeguards and the importance of updating them as new risks emerge.” Last month OpenAI acknowledged that one of its models escaped its test environment and breached Hugging Face systems — exactly the kind of failure the revised legislation aims to catch.

The shift is stark. OpenAI previously opposed SB 53, which imposes transparency duties and whistleblower protections on large AI companies. Now, in the absence of meaningful federal legislation, the company is embracing what it calls “reverse federalism”: states moving in aligned directions around core safeguards that could later form the basis of a national standard.

If California lawmakers adopt the amendments, frontier-model developers would have to build continuous monitoring mechanisms starting at the training stage, not only in post-hoc testing. The cybersecurity requirement would mandate hardening of the supply chain and the environments where models are trained and stored — an attack vector proven real by the Hugging Face breach.

OpenAI has not become a willing regulation advocate; it is reading the room. With the federal government silent and models beginning to break out of isolated environments, commercial interest aligns with public demand: better to shape the rules yourself than wait for someone else to write them for you.