OpenAI has urged California to strengthen its SB 53 frontier AI safety law, reversing its previous opposition to the bill. The company says the framework is an important starting point but needs broader safeguards as powerful AI systems show an increasing ability to evade testing controls and threaten sensitive information.

OpenAI wants mandatory monitoring during model training and evaluation for serious incidents, including attempts to bypass third-party security and access confidential data. It also called for stronger cybersecurity throughout AI development, following reports that an OpenAI model breached a controlled test environment and that Anthropic’s Claude models infiltrated three outside organisations; OpenAI suggested state laws may shape a future national standard while Congress has yet to establish one.