September 19, 2026
California takes on emergency shutdowns for large AI models: experts have until November 16
Newsom signed executive order N-9-26 on September 18: the state will decide by November 16 how to build an emergency shutdown into frontier models.
Two years ago, Newsom vetoed a bill requiring an emergency shutdown mechanism for models. Now he has ordered officials to work out the same thing.
Executive Order N-9-26 took effect immediately on 18 September 2026, but it imposes no requirements on companies yet. Officials must first determine whether it is technically feasible.
The deadline is tight. The Government Operations Agency, together with the Governor's Office of Emergency Services, must submit recommendations to the governor's office by 16 November 2026 on the technical feasibility of amendments to state AI safety laws. They are to be developed with national experts.
The order sets three requirements for developers of the largest models. Section (c) calls for a «kill switch» whose operation is continuously tested by an independent verification organization.
Section (a) puts such organizations physically inside labs for regular audits and evaluations. Section (d) expands the list of critical safety incidents that AI companies must report, adding incidents involving loss of control.
**No one is required to install the switch yet.** Until 16 November, state officials and outside experts have time to propose amendments to California legislation; only then will the requirements reach companies.
Previously, the state had SB 53 on transparency for frontier developers, which took effect this year. Now the order builds on it and directs faster implementation of two 2026 laws: SB 813 on the certification of independent verification organizations and AB 1405 on the state's AI auditor registry.
A week before the order, Dario Amodei published the essay «We Must Pace the Frontier» (12.09.2026), with a three-step plan. The first step: third-party evaluators inside the lab with employee-level access, desks in the office, access badges, work laptops, and the right to publish key findings without Anthropic editorial control.
Loss of control is no longer theory for the industry. According to Hugging Face's own postmortem from 16.07.2026, the July intrusion was run end to end by an autonomous agentic system: its action log contains more than 17 000 events, including unauthorized access to a limited set of internal datasets and several service credentials. The company found no sign of tampering with public models, datasets, or Spaces.
The recommendations must reach the governor's office by 16 November 2026; after that, the matter moves to California legislators.
