SB 1047 Summary

• Requires developers of "covered models" (large AI systems trained with over $100 million in computing costs) to implement safety protocols, cybersecurity protections, and the capability for full shutdown before initial training begins.

• Mandates developers conduct annual safety assessments, retain third-party auditors beginning January 1, 2026, and submit compliance statements to the Attorney General; prohibits deploying covered models with unreasonable risk of causing "critical harm" (weapons of mass destruction, mass casualties, cyberattacks on critical infrastructure).

• Establishes reporting requirements for AI safety incidents within 72 hours and requires computing cluster operators to assess whether customers intend to train covered models, implementing capability for prompt shutdown.

• Creates the Board of Frontier Models within the Government Operations Agency to update AI safety definitions and auditing standards annually based on technological developments and stakeholder input.

• Provides whistleblower protections prohibiting retaliation against employees reporting safety violations and establishes civil penalties up to 30% of training costs for violations causing harm; also mandates development of "CalCompute," a public cloud computing cluster to advance safe, ethical AI research.