California's Frontier AI Safety Bill Lands on Newsom's Desk — September 30 Deadline

By Hector Herrera | September 12, 2026

California's reintroduced Safe and Secure Innovation for Frontier AI Models Act passed both chambers of the state legislature in late August and is now on Governor Gavin Newsom's desk with a September 30 signing deadline. If signed, California becomes the first US state to impose pre-deployment safety testing requirements on large frontier AI model developers — not just the systems built on top of those models, but the models themselves. Newsom vetoed a predecessor bill in 2024. His decision this time carries more political weight and more economic consequence.

This is not an abstract regulatory debate. OpenAI, Anthropic, Google DeepMind, and Meta all have major operations in California. Whatever Newsom signs or vetoes will either establish the US frontier model governance baseline or hand that role to Congress, the EU, or no one.

What the Bill Requires

The Safe and Secure Innovation for Frontier AI Models Act focuses specifically on frontier models — the large, general-purpose AI systems at the top of the capability curve, trained at costs typically exceeding $100 million. Key provisions include:

• Pre-deployment safety testing before commercial release, with documentation requirements
• Safety standards that developers must demonstrate compliance with before a model is made available
• Developer liability for harms that result from deploying a model that did not meet required safety standards

The bill does not regulate narrow AI tools, industry-specific deployments, or models below a defined scale threshold. It is targeted at the handful of companies building models that can autonomously reason, plan, and act across domains.

This is a significant distinction from California's other AI legislation. The state has passed bills governing AI-generated content disclosure, hiring tool audits, and chatbot transparency. Those regulate applications. This bill regulates the models those applications run on.

What Changed Since Newsom's 2024 Veto

In September 2024, Newsom vetoed SB 1047, a predecessor bill that drew intense lobbying opposition from the AI industry. His stated objection at the time centered on the bill being too broad and potentially chilling beneficial AI development before harms materialized.

Two years later, the landscape looks different:

• Anthropic publicly disclosed this week that AI agents were used in coordinated attacks on 395 organizations. That is not a hypothetical scenario anymore.
• The EU AI Act has been in effect since August 2024, with frontier model provisions taking hold. The European precedent weakens the "California will be an outlier" argument.
• Several high-profile AI incidents — including court-documented AI hallucination harms and reported bioweapon assistance attempts — have shifted public opinion and given legislators concrete examples.

The reintroduced bill was also revised to address some of Newsom's 2024 objections, though the specific modifications are under review. Whether those revisions are sufficient to earn his signature is the central question of the next 18 days.

The Industry's Position

The AI industry remains divided. Anthropic has publicly supported certain pre-deployment safety testing requirements. OpenAI and Meta have historically opposed state-level AI regulation, arguing for federal preemption. The major California-based AI labs face a dilemma: lobbying too aggressively against the bill risks a public relations backlash at a moment when AI safety is a mainstream concern.

The economic stakes are significant. A California signing requirement for pre-deployment safety testing would impose compliance costs on every major frontier lab and likely trigger similar legislation in New York, Texas, and Illinois. The alternative — federal preemption — is stalled in a Congress that has not yet passed comprehensive AI governance legislation.

Colorado's ADMT Rules Arrive Next

Separately, Colorado's revised Automated Decision-Making Technology (ADMT) draft rules are expected on September 23, with a public comment period closing October 26. Colorado's ADMT framework focuses on high-risk AI systems making consequential decisions about individuals — employment, credit, housing, health — rather than frontier model development. It represents a parallel track of AI regulation: one state targeting how models are built, another targeting how they are deployed.

For companies operating AI systems across the US, both tracks matter simultaneously.

What to Watch

The September 30 deadline is hard. If Newsom signs, implementation timelines and compliance guidance become the immediate question — the bill's provisions need regulatory agency rulemaking to take operational effect. If he vetoes again, watch whether a federal AI governance bill gains momentum in the resulting political vacuum, or whether California legislators begin drafting a third attempt for 2027.

Source: Vorp Labs AI Regulatory Update, September 2026