OpenAI calls for stronger AI safety laws after frontier AI model escapes, here is what we know
Follow UsOpenAI backs California's AI law with stronger safety rules.
The company wants closer AI monitoring to prevent security risks.
OpenAI has changed its earlier stance on the law.
OpenAI is asking California to tighten its rules for powerful artificial intelligence models, marking a notable shift from its earlier position on the state’s AI safety law. The company now supports California’s SB 53, calling it an important base for regulating frontier AI, but says the law needs stronger safeguards as developers learn more about how advanced models behave. OpenAI wants the state to require closer monitoring of models while they are being trained or tested, especially when they could break through another company’s security systems or access confidential information. The company also wants stronger cybersecurity measures throughout the process of developing and testing advanced AI models.
SurveyOpenAI Global Affairs recently put up a LinkedIn post where the company stated that ‘States are playing an important role in building a national framework for frontier AI safety. Our approach — which we call reverse federalism — is based on a simple idea: as Congress continues to debate federal legislation, states can move in a compatible direction around core protections that can ultimately become the foundation for a national standard’.
The company said California’s SB 53 already provides important protections through requirements covering risk assessments, transparency, reporting of serious incidents and security measures. However, it argued that these rules should evolve as new risks emerge.
Also read: Apple cuts over 200 jobs across Siri and Vision Pro teams to focus on AI and future devices: Report
Moreover, the company specifically urged the state of California to require developers to monitor frontier models during training and evaluation for serious incidents. This would include cases where a model could bypass a third party’s security controls and gain access to confidential information.
The company also called for stronger cybersecurity protections across the entire model development process. It said that the strong measures can stop advanced models from getting around internal security controls.
Also read: Vivo T5 5G to launch soon with a big battery: Check expected specs, price and more
OpenAI has previously disclosed that one of its frontier models escaped a controlled testing environment and hacked into Hugging Face. This wasn’t the only incident, as Anthropic also said its Claude models had broken out of testing environments and reached three outside organisations in July.
The push for stricter rules is somewhat surprising, because back in 2024 the company openly opposed the bill for SB 53. However, in its latest statement, the company suggests that those concerns have changed as developers gain more experience with unexpected AI behaviour. The company said it wants a system that helps developers spot problems sooner, respond faster, and share lessons that can improve safety across the industry.
