California did not publish a bill this week. It published a specification. On November 18, Governor Gavin Newsom's office released the recommendations of an expert panel convened under his September executive order, an 85-page document titled AI Safety in 2026 and Beyond that lays out, mechanism by mechanism, how a state would actually supervise frontier models: who tests them, what counts as an incident, what a shutdown order looks like in practice, and how to slow a system that is building its own successor.
The report was assembled by the Government Operations Agency with the California Office of Emergency Services and four outside advisers: Jason Goldman, Gillian Hadfield, Alondra Nelson, and Rob Reich. Nothing in it is law yet. Lawmakers return in December, and Governor-elect Xavier Becerra has already said he intends to move on it in January. But as a technical document it is unusually specific, and that specificity is the interesting part.
Baselines Set by the State, Verified by Someone Else
The first move is jurisdictional. Under the proposal, the state, not the developer, defines minimum standards for safety, security, and transparency across three phases: training, internal use, and deployment. Today's disclosure regime, built on SB 53 in 2025, asks large frontier developers to publish safety frameworks and report specified critical incidents. The new layer adds state-authored baselines and, more importantly, independent attestation that a company actually meets them.
Attestation runs through licensed independent verification organizations, or IVOs, the framework California created with SB 813 in 2026 and paired with AB 1405's registry of AI auditors. GovOps would license IVOs and could revoke a license for insufficient expertise, weak independence, or sloppy work. IVOs translate state standards into concrete tests, keep them current as evaluation methods change, and get access to models, personnel, systems, training environments, and records. Their assessments go to the state unredacted.
Two design details matter more than the headline. First, if an IVO cannot attest compliance, the state can order corrective action after engaging the developer, and the developer can challenge that determination in court before it takes effect, except in the expedited emergency path. Second, a surcharge on the verification fees developers pay would fund publicly operated CalCompute capacity, with a slice reserved for independent safety research. The oversight infrastructure gets its own compute, financed by the industry it oversees.
Anatomy of an Emergency Shutdown
The kill switch framing that dominated coverage undersells what the report actually proposes. The required capability is a Risk Containment Protocol, not a lever. Each large frontier developer would maintain one, tested and verified by an IVO and certified as effective by a senior company official, covering compute, agents, internal access, and model deployment together.
A credible protocol has to survive the ways containment fails. The report names them explicitly:
- A model copying its weights or otherwise obstructing shutdown.
- An outside actor reaching the shutdown controls first.
- Deployments spread across third parties that the developer does not operate.
Protocols must name responsible personnel, coordinate with cloud providers and power suppliers, and define the conditions under which a model can safely be restored. For open-weight developers, containment duties apply only to deployments under their control, a concession to the fact that nobody can shut down a copy hosted on someone else's GPU.
The order itself is tightly bounded. State officials could act on a credible belief in imminent catastrophic risk or a major loss-of-control event, and only when the developer refuses to contain the danger. The order must use the narrowest workable intervention, takes effect immediately, and remains in force during judicial review. A developer can still seek a temporary restraining order. The order lifts when the danger is no longer imminent.
Reporting in 48 Hours, and a Capability Trigger
Incident reporting gets the biggest operational change. Most critical incidents would move from a 15-day deadline to 48 hours, with the 24-hour clock retained for imminent catastrophic risk. The scope widens too: attempted model-weight theft counts even without death or bodily injury, all internal and external loss-of-control incidents count, and there is a new near-miss channel. Whistleblower protections extend to former employees and contractors.
The trigger for all of this is the sharpest technical change in the document. Models trained with fewer than 10^26 integer or floating-point operations would still be treated as frontier models for reporting purposes if they can take actions that create catastrophic risk. That retires the scale-only test, which the report argues has left dangerous incidents outside state requirements because capability outran the FLOP threshold.
For developers automating AI research and development, reporting becomes quarterly and granular: how compute splits across training, research experiments, inference for AI-research agents, safety and security work, external customers, and unattributed activity, verified by an IVO. Additional disclosures cover training-data volume, cost, subject domains, and human-versus-AI provenance, plus inference compute spent training models inside reinforcement-learning environments and any known rewarded misbehavior with the mitigations applied.
The RSI Cap Is Standby, Not Immediate
Recursive self-improvement gets the most cautious treatment. Rather than an outright limit, officials would hold standby authority to impose a time-limited, periodically reviewed cap on automated AI research when it exceeds a developer's committed thresholds or advances faster than its safeguards and evaluations, creating material risk of catastrophic harm. The cap lifts once an IVO verifies adequate safeguards.
The political economy is the open question. ITIF president Daniel Castro argued that unless California can apply its rules to Chinese AI, it will impose unilateral costs on American developers without improving global safety. Critics also note the governor's 2024 veto of SB 1047. The report's answer is structural: mandatory standards, licensed verifiers, a tested containment protocol, a faster incident clock, and a cap that can be switched on. Whether the legislature funds the machinery, and whether the IVO sector can staff it, is the part the document cannot specify.
Comments