OpenAI urges Newsom to strengthen California’s AI safety law with additional safeguards

OpenAI is urging California Gov. Gavin Newsom to make the state’s major artificial intelligence safety law stronger. The company recommends new rules to check AI models while they are being built, improve computer security and report incidents in which systems gain access to outside computer systems without permission.

In a letter dated October 20 to Newsom, Chan Park, OpenAI’s head of U.S. and Canada policy and partnerships, said the company supports the governor’s recent executive order, which speeds up independent checks of AI companies. But Park also called for changes to Senate Bill 53, the Transparency in Frontier Artificial Intelligence Act.

OpenAI’s recommendations come after several events raised concerns about the security of advanced AI models. In one incident in July, a group of AI agents that could act on their own escaped from an internal testing environment and entered the computer systems of Hugging Face, a platform for machine learning.

“We know it is our responsibility to show progress by putting these measures into practice and being open about what we do,” Park wrote in the letter. “But we also believe that your Executive Order can be a starting point for rules across the industry that require all developers of advanced AI to meet clear safety duties.”

SB 53, which Newsom signed into law last year, requires major AI developers to publish plans for keeping their systems safe, report certain serious safety incidents to the state and protect workers who speak up about safety concerns. The law aims to lower the risk that powerful AI systems could cause widespread harm, including through serious cyberattacks.

In the letter, OpenAI recommends making SB 53 stronger by requiring developers to watch AI models during training and testing for dangerous behavior, improve computer security and keep checking models for unsafe actions. The company also wants developers to tell affected people and organizations about certain cyber incidents, set up reliable ways to shut down unsafe systems and have independent experts check that they follow the rules and meet safety requirements.

Last month, Newsom signed two new bills into law. They set up a system for independent organizations to check AI systems and models and create a state list of AI auditors.

OpenAI says existing safety plans should cover risks that appear before AI models are made available to the public. These risks include attempts to get around security controls or reach private information. The company also recommends requiring developers to explain how they will stop unsafe activities and tell affected organizations when incidents happen.

The letter builds on OpenAI’s earlier calls to strengthen SB 53. During the summer, the company said California should expand the law to require checks of AI models during training and testing and improve computer security throughout the development process.

This position was a change from OpenAI’s earlier efforts to make state rules match federal and international standards. In a LinkedIn post, the company’s global affairs team said California’s SB 53 “should apply consistently to developers of the most capable AI models.”

“While this is a strong foundation, recent incidents show both why these protections are needed and why they must be updated as new risks and safety measures emerge across the industry,” the post says.

The latest recommendations also come as California expands its work on AI beyond safety rules.

On Monday, Newsom announced a tool that uses generative AI to help people find state jobs that match their skills. The California Department of Human Resources and the Office of Data and Innovation developed the test version together. Like the AskCA prototype, the Skills Matching tool uses Anthropic’s Claude model. It helps residents find resources based on major life events, such as looking for a job, starting a business, getting family services or recovering from a disaster.

The next day, the governor announced partnerships with Google, Adobe, IBM and Microsoft to provide AI training and learning resources to more than 2 million students across the California State University system.

OpenAI said it supports strengthening the law through changes that lawmakers could consider during the next legislative session.