California's Newsom Orders a 60-Day Study on an AI 'Kill Switch' for Frontier Models

California Governor Gavin Newsom signed Executive Order N-9-26 on September 18, 2026, directing the state's Government Operations Agency to complete a 60-day study — due November 16 — into whether California should require a mandatory emergency shutoff mechanism for frontier AI models that "go rogue," alongside a parallel study on placing independent verification auditors onsite at large frontier labs. It's a study order, not a binding rule yet, but it's the most concrete state-level move toward a formal AI "kill switch" requirement to date.

Update — September 20, 2026: As reporting on the 60-day deadline circulates more widely, two questions keep coming up that the initial coverage under-addressed: whether this is a legal mandate on labs or a technical spec, and whether it would touch API access for downstream builders or only frontier-scale training runs. We've expanded the sections below — "What the study process itself will likely have to grapple with" and "What this means for builders" — to answer both directly, and added the SB 1047 precedent that explains why California is the state driving this rather than any other.

TL;DR

What the order actually directs

Executive Order N-9-26 is specifically a study order, not an immediate regulatory mandate — an important distinction given how "AI kill switch" headlines can read as if California has already imposed a new binding requirement on AI companies. What Newsom actually signed directs California's Government Operations Agency to spend the next 60 days researching and preparing recommendations on two connected but distinct questions: first, whether the state should require frontier AI developers to build in a mandatory emergency shutoff mechanism — a way to forcibly halt a model's operation if it begins behaving in unintended, dangerous, or "rogue" ways; and second, whether the state should place independent verification auditors physically onsite at large frontier AI labs to verify safety claims directly rather than relying on labs' own self-attestation. The study is due November 16, 2026.

The onsite-auditor idea echoes a trend already underway voluntarily

The second half of this order — independent auditors embedded directly inside frontier labs — isn't a novel concept invented here; it's the same underlying idea several labs have already begun adopting voluntarily. explainx.ai covered Anthropic's "We Must Pace the Frontier" essay, which explicitly commits Anthropic to giving embedded third-party evaluators employee-level access, and its $1 billion partnership with Accenture building out exactly this kind of embedded-evaluation infrastructure. What Newsom's order signals is California potentially moving that same idea from a voluntary, lab-chosen commitment to a state-mandated requirement applying across the industry — a meaningfully different posture, since voluntary commitments can be scaled back at a company's discretion while a regulatory requirement would not carry that same flexibility.

How this differs from the federal kill-switch effort

It's worth being precise about which "AI kill switch" story this is, since there's more than one currently in play. explainx.ai previously covered Sen. Rand Paul blocking a federal AI kill-switch bill in the US Senate — that was a proposed piece of federal legislation that stalled. Newsom's order is a separate, state-level executive action, and it isn't dependent on or blocked by the federal bill's status. This is one of the clearer recent examples of the broader pattern in 2026 AI regulation: as federal legislative efforts stall or move slowly, individual states are increasingly pursuing their own AI regulatory frameworks independently, and California — home to the largest concentration of frontier AI labs — is positioned to have outsized influence on what those state-level rules eventually look like in practice.

Why California specifically has outsized leverage here

It's worth naming directly why an executive order from one US state carries more practical weight in AI regulation than the same order coming from most other states would. California hosts the largest concentration of frontier AI lab headquarters and infrastructure in the country — OpenAI, Anthropic, Google DeepMind, and Meta's AI divisions all maintain substantial California operations, meaning California-specific regulatory requirements can, in practice, reach a meaningful share of the entire frontier AI industry without needing federal legislation at all. That's structurally different from a similar order issued by a state with little AI industry presence, where compliance would be a narrower, more easily worked-around obligation. It's also part of a broader, recognizable pattern in US tech regulation history — California's own environmental and privacy regulations (CCPA being the clearest recent example) have repeatedly ended up functioning as de facto national standards, because it's often more efficient for a company to build one compliant product or process than to maintain a California-specific version alongside a separate national baseline. If Newsom's kill-switch and onsite-auditor study eventually produces binding requirements, this same dynamic is a reasonable base rate to expect: California-specific AI safety rules have a real chance of becoming the practical national standard, regardless of what happens at the federal level.

This is also not California's first attempt at frontier-model-specific safety legislation, which is part of why the state has both the institutional muscle and the political scar tissue to move quickly here. Newsom vetoed SB 1047 — the comprehensive frontier-model safety bill that would have required pre-deployment safety testing and a state-enforceable shutdown capability for the largest models — in 2024, citing concerns that it regulated based on model size (compute thresholds) rather than actual risk of harm. That veto didn't end the effort; it rerouted it. The legislature came back with the narrower SB 53 framework, and most recently SB 813 and AB 1405, which explainx.ai covered in detail — those two bills don't mandate audits themselves, but they license and register who is legally allowed to perform an AI audit once some other California law requires one, laying exactly the kind of verification infrastructure a kill-switch mandate would eventually need to lean on. Executive Order N-9-26 reads as the next link in that same chain: SB 1047's ambition, SB 53's narrower scope, SB 813/AB 1405's auditor infrastructure, and now a dedicated study on the one mechanism SB 1047 never fully resolved — a mandatory emergency shutoff.

The politics of "kill switch" framing itself

It's also worth flagging how loaded the specific term "kill switch" is doing work in how this story is being covered and discussed, separate from the substance of the order itself. "Kill switch" evokes a dramatic, single-button emergency-stop mechanism, which makes for a punchier headline than the more accurate but drier description of what's actually been ordered — a 60-day interagency study process that may eventually recommend some form of mandatory emergency shutoff capability. That gap between the dramatic framing and the more incremental regulatory reality underneath it is common across AI policy coverage generally, and it's worth keeping in mind when evaluating how urgently or dramatically to react to any single AI regulation headline: the underlying process is almost always slower, more contingent, and more subject to negotiation and revision than the initial framing suggests.

What the study process itself will likely have to grapple with

Sixty days is a genuinely tight timeline for the Government Operations Agency to work through the actual hard design questions a mandatory emergency shutoff mechanism raises — questions that go well beyond simply deciding a kill switch should exist. Technically, an effective shutoff mechanism for a deployed, distributed frontier model isn't a single physical button; it likely requires some combination of API-level access revocation, infrastructure-level compute deprovisioning, and coordination with whatever cloud or on-premises infrastructure the model runs on, each of which raises its own implementation and jurisdiction questions when the underlying infrastructure spans multiple states or countries. There's also a harder definitional question underneath the technical one: what specifically counts as a model having "gone rogue" in a way that triggers the mechanism, who has the authority to make that determination in a time-sensitive situation, and how the state avoids either false positives (shutting down a model over an ambiguous or disputed behavior) or false negatives (a genuinely dangerous behavior that doesn't clearly meet whatever threshold gets defined). None of these are unanswerable questions, but they're the kind of question a 60-day study process typically produces a framework for addressing rather than a fully resolved, ready-to-implement answer to — worth setting expectations accordingly for what the November 16 study output is likely to actually contain.

Is this a legal requirement on labs, or a technical mandate?

It's worth being precise here, because the order's language leaves this genuinely open. As written, N-9-26 orders a study into whether to impose a legal requirement — it does not itself specify a technical standard a shutoff mechanism must meet. If the study's recommendations become law, the most likely shape (based on how SB 1047, SB 53, and SB 813/AB 1405 were all structured) is a legal obligation stated as an outcome — "developers of covered frontier models must maintain the capability to disable or restrict a model's operation within [some time window] of a qualifying trigger" — rather than a prescriptive technical spec dictating exactly how that capability must be built. That distinction matters enormously for compliance cost: an outcome-based requirement lets a lab satisfy it however fits its own infrastructure (API kill-switches, canary deployments it can roll back, compute-level throttling), while a prescriptive technical mandate would force every covered lab onto the same architecture regardless of fit.

The genuinely hard engineering problem underneath the political framing

Set the politics aside for a moment and the honest engineering reality is this: "kill switch" implies a single, instantaneous action, but shutting down a frontier model already serving production traffic is closer to a distributed-systems incident-response problem than a button press. A model like GPT-5-class or Claude-class systems can be answering millions of API calls already in flight at the moment a shutdown is triggered — mid-generation completions, agentic workflows that have already taken real-world actions (sent emails, executed trades, written code that's queued for deployment), and downstream applications built on top that have no warning a cutoff is coming. A responsible shutdown mechanism has to answer questions a simple on/off switch doesn't: Do in-flight requests get allowed to complete or hard-terminated mid-token? Does the cutoff apply globally and instantly, or does it need a graceful drain period that a "rogue" model scenario may not have time for? Who is liable for downstream failures at companies that had no notice their dependency just vanished? None of this makes a shutdown capability impossible to build — most labs already have some version of a kill mechanism for security incidents — but it does mean a 60-day study realistically produces a governance framework and a set of outcome requirements, not a finished technical specification ready for engineers to implement on day one.

Honest limitations

• This is a study order, not a final policy — the specific design of any eventual kill-switch requirement (what counts as "rogue" behavior, how the shutoff mechanism would technically work, which companies and model sizes it would cover) is explicitly left to the 60-day study process, not specified in the order itself.
• No enforcement mechanism or penalty structure exists yet — because nothing is mandatory at this stage, there's nothing yet to enforce.
• Which specific companies would ultimately be covered isn't defined in currently available order text — "large frontier AI labs" is descriptive, not a named or thresholded list.
• The study's actual recommendations (due November 16, 2026) haven't been published yet — this post covers what was ordered, not what the study will ultimately recommend.

What this means for builders

If your company operates a frontier-scale AI model with any California nexus, this is worth tracking closely over the next two months — the November 16 study deadline is the concrete date to watch for what a mandatory kill-switch and onsite-auditor framework might actually require in practice. For teams at smaller AI companies not in the "large frontier lab" category this order is aimed at, the more useful signal is directional: California is actively building the regulatory infrastructure and political appetite for mandatory AI safety verification, and that same infrastructure — onsite auditors, emergency shutoff requirements — has historically had a way of expanding in scope over subsequent legislative cycles once an initial framework exists, so it's worth watching even if your company isn't currently in the order's stated target range.

Compliance timeline to actually track: the only hard date right now is November 16, 2026, when the Government Operations Agency's study is due. Nothing is binding before then — there is no rule to comply with today. If the study recommends legislation, expect a bill introduction in the following legislative session, then a normal bill process (committee, floor votes, potential veto) before anything becomes enforceable, likely pushing real compliance obligations well into 2027 at the earliest, consistent with how SB 813 and AB 1405's own deadlines were staggered out to 2028 and 2029.

Does this touch API access, or only frontier training runs? Based on the order's own framing and every prior California AI bill in this lineage (SB 1047, SB 53, SB 813/AB 1405), the target is consistently frontier model developers — companies training and deploying the largest, most capable models above some compute or capability threshold — not the far larger population of companies calling those models' APIs to build products. If you're building an app on top of OpenAI, Anthropic, or Google's APIs rather than training your own frontier-scale model, you are very unlikely to be a "covered developer" under any version of this that has been proposed so far; the obligation, if it materializes, would sit with the labs whose APIs you call, not with you. The one indirect exposure worth planning for: if a lab-level kill switch is ever triggered against a model you depend on, your own product inherits that outage with effectively zero notice — which is a good independent reason to build fallback model providers into any production AI feature regardless of what California ultimately requires of the labs themselves.

Related on explainx.ai

• California signs SB 813 and AB 1405 — the AI auditor licensing precedent behind this order
• Trump's poll to rename AI "Superior," "Extreme," or "Supreme" Intelligence — one day after this order
• Sen. Rand Paul blocks a federal AI kill-switch bill
• Hawley and Blumenthal demand a floor vote on the FRONTIER AI Act
• Dario Amodei: "We Must Pace the Frontier" — embedded evaluators
• Anthropic and Accenture partner on embedded AI evaluation
• AI regulation: EU AI Act and US policy, a complete guide
• OpenAI discloses 6 model safety incidents, warns against max-speed scaling
• Primary sources: gov.ca.gov official release · Bloomberg

This post is sourced to California's official gov.ca.gov release of Executive Order N-9-26 and Bloomberg's coverage, both dated September 18, 2026. The order's 60-day study is due November 16, 2026 — check for updated coverage once that study is published.