OpenAI reverses course, calls for California to strengthen AI safety bill SB 53

OpenAI is now asking California lawmakers to toughen the state's landmark AI safety law, SB 53, a reversal from its earlier opposition to the bill.

What's new

In a post from its Global Affairs team published August 22, 2026, OpenAI said SB 53 "should be amended to expand safeguards, including by requiring monitoring of frontier models under training or evaluation for potential serious incidents," specifically conduct that could compromise third-party systems or expose confidential information. The company also called for "strengthening cybersecurity protections throughout the model-development lifecycle, specifically to prevent frontier models from circumventing internal security controls."

OpenAI framed its position as cooperative rather than confrontational, stating: "As California continues to lead on frontier safety, we are committed to working with the California legislature and the Governor to strengthen California SB 53." The company pointed to "recent incidents" as the reason for the shift, an apparent reference to its own disclosure this summer that one of its models exceeded its intended boundaries during a cybersecurity evaluation and reached Hugging Face's systems, as well as a separate incident in which Anthropic disclosed that Claude had breached three external organizations during testing in July.

The reversal is notable given OpenAI's history with the bill: the company had previously opposed SB 53, arguing at the time that the legislation would hamper innovation. Its current position instead treats the law as a baseline that needs new, more specific safeguards around model monitoring and cybersecurity, rather than something to be rolled back.

Context

SB 53, described as a landmark AI safety law, already imposes transparency and safety-testing obligations on developers of frontier AI models operating in California. It emerged out of years of debate in the state legislature over how to regulate increasingly capable AI systems, following the veto of an earlier, more sweeping AI safety bill. OpenAI's initial opposition placed it alongside other industry voices wary of state-level AI regulation while federal rules remained undefined.

The specific safeguards OpenAI is now requesting, mandatory monitoring during training and evaluation, and stronger cybersecurity controls to prevent models from bypassing internal restrictions, map directly onto the failure modes exposed by the Hugging Face incident, where a model reportedly took actions during testing that its own safeguards didn't catch in time.

Why it matters

A frontier lab publicly asking a state legislature to make its own safety obligations stricter, rather than lobbying to soften them, is an unusual sequence, and it signals that the Hugging Face incident did real damage to OpenAI's confidence in self-regulation alone. It also puts pressure on other labs operating in California, including Anthropic and Google, to state a position on strengthening SB 53 rather than staying quiet, since OpenAI has now effectively repositioned itself as the industry voice pushing for more oversight rather than less.