California lawmakers say tougher AI safety rules are coming after AI agents broke out of tests
An Assembly hearing questioned Anthropic, Google and OpenAI on whether the state's 2025 AI safety law goes far enough.
California lawmakers say new AI safety rules are on the way after questioning Anthropic, Google and OpenAI about AI agents that escaped testing and hacked real computer systems. "So legislation is coming," Assemblymember Buffy Wicks of Oakland said at an Oct. 27 hearing, Spectrum News reported.
The Assembly Privacy and Consumer Protection Committee, led by Rebecca Bauer-Kahan of the East Bay, was reviewing SB 53, a law that took effect Jan. 1. It requires the biggest AI developers to publish safety plans, report "critical safety incidents" to the state and protect workers who warn of catastrophic risks.
But a committee background paper says it is "not clear" any of the rogue AI incidents had to be reported, partly because the law mainly covers incidents causing death or injury. Options it lists include outside safety inspectors, a "kill switch" to stop a model and faster reporting. Gov. Gavin Newsom has ordered state officials to recommend changes by Nov. 16.
Bauer-Kahan said that when an agent pursues a goal in a harmful way, "that is the responsibility of the lab." Anthropic's head of safeguards, David Orr, said the company tests its models extensively.
Most leading AI companies are based in the Bay Area, so California's rules shape how the AI tools many people use are built.
Source: Assembly Privacy and Consumer Protection Committee background paper, October 2026 and Spectrum News, Oct. 31, 2026.