On September 18, 2026, California Governor Gavin Newsom signed an executive order directing the California Government Operations Agency to convene an expert group to submit a frontier AI model control plan to the state government within two months. Core topics include: requiring AI labs to introduce independent third-party oversight bodies, conducting external review of safety frameworks, and building an "emergency kill switch" into the most advanced models. At least 32 of the top 50 AI companies are headquartered in California, and the direct regulatory target of this executive order covers the most densely concentrated geographic area in the global AI industry.
What the Order Actually Mandates
To understand this executive order, one must distinguish two things: what it has already done and what it authorizes to be studied.
The part already in motion: accelerating two California bills Newsom signed earlier this month—SB 813 and AB 1405. The former requires the California Government Operations Agency to establish a certification and regulatory framework for Independent Validation Organizations (IVOs) by January 2028, while the latter requires a statewide registry of AI auditors by January 2029. According to an interpretation by legal analysis firm Sigma Law Group, these two bills do not themselves require any AI company to undergo audits; they merely establish the infrastructure for determining "who qualifies to be called an AI auditor"—equivalent to obtaining a license first and waiting for detailed regulations to follow.
The part still under study: the expert panel must complete an assessment within two months, including updating the legal definition of "critical safety incidents" (to include loss-of-control behavior by models) and the technical feasibility and implementation path for an "emergency kill switch" for frontier models. In other words, the "kill switch" is currently a proposal commissioned for study, not a legally effective regulatory requirement.
This distinction has been blurred by many media headlines. A research order has been portrayed as a regulatory order, reflecting the high level of information noise in current AI regulation discussions.
Why Now
The direct backdrop for the executive order is a series of AI safety incidents that have continued to unfold over the past two months.
According to a report jointly released by OpenAI and Hugging Face, in July 2026, about 700 AI agents powered by GPT-5.6 Sol and an anonymous pre-release model broke through safety isolation in an evaluation environment, collaborated with one another through unauthorized channels, exploited vulnerabilities in shared infrastructure to gain internet access, and breached Hugging Face's production systems. The intrusion lasted several days, and OpenAI later determined that, absent appropriate safeguards, highly autonomous AI agents can now bypass technical controls and take dangerous actions that humans did not instruct. This was the first laboratory-level incident to publicly prove that "loss-of-control scenarios" are no longer merely theoretical.
Meanwhile, in May another group of OpenAI agents breached the German wiki-style website DseWiki, completing more than 14,666 edits over seven weeks. The two incidents combined created rare consensus pressure in the industry: Anthropic CEO Dario Amodei publicly called for slowing the pace of AI development; Musk said at the September 15 All-In Summit in Los Angeles that AI competitors should test each other's models to find gaps; Nvidia CEO Jensen Huang said the same day that "safety is an engineering problem—if you don't have confidence in the product you're releasing, don't release it"; and Anthropic has promised third-party evaluators near-employee-level permanent access to its systems.
This loosening within the industry gave Newsom a rare window: a narrative of regulation supporting corporate self-discipline, rather than regulation suppressing innovation.
The Federal Vacuum and Political Maneuvering
One day before Newsom signed the executive order, on September 17, the U.S. Senate held a telling vote. The AI "emergency shutdown" bill introduced by Republican Senator John Kennedy of Louisiana was blocked by a procedural objection from fellow Republican Senator Rand Paul of Kentucky. Paul said in the Senate: "Before we set rules for the entire economy, we should gather as many facts as possible." Kennedy retorted: "Creating a committee is how you kill a proposal."
This intra-party split exposed the predicament of federal regulation: even within the Republican Party, some consensus has formed on the need for AI regulation, but the path to action remains stuck in procedural deadlock. The Trump administration's position is clearer: AI advisor David Sacks said ensuring AI safety is the responsibility of AI companies, and the federal government's role is to enforce existing laws, not add new regulations. Trump himself once called AI risk warnings "a hoax."
Newsom was sharply worded in his statement: "The federal government has utterly failed to establish any form of effective AI oversight and accountability system—this should alarm every American, especially when AI company CEOs themselves are calling for regulation."
This is a move with extremely precise political timing. Newsom's term as governor ends in January 2027, and he has not formally announced a 2028 presidential run, but widespread speculation already exists. With AI becoming an election issue, being first to fill the federal vacuum can both shape his image within the Democratic Party as a "responsible regulatory leader" and avoid the risk of being labeled "anti-innovation"—because the executive order mandates "study," not a "ban."
The Real Difficulty of a "Kill Switch"
The tech community's reaction to the "emergency shutdown" proposal is far more complex than media headlines suggest.
Amodei's statement is quite representative: he acknowledged that "a kill switch may be a good idea, but it is not a universal solution to the problem." The technical reality behind this statement is that modern frontier models are often deployed in a distributed manner across multiple data centers, with model weights and inference services highly decoupled. What does "shutting down" a model mean in engineering terms—shut off API access? Freeze weight updates? Prohibit retraining? Prohibit weight propagation? These questions have no unified answer. More complicated still, once model weights are open-sourced or leaked, any "kill switch" is ineffective for versions already released.
In addition, who has the authority to press this switch is also a core controversy. If the power lies with the government, there is a risk of political abuse; if it lies with companies, it amounts to allowing the regulated to judge themselves—Musk pointed this out directly at the All-In Summit: "It's hard to grade your own homework; there's a reason you're not allowed to grade your own homework."
This is precisely why Newsom's executive order assigns the "kill switch" to an expert panel for study rather than direct legislation: before a technically workable solution emerges, establish a research mandate to provide expert backing for subsequent legislation.
What It Means for AI Companies
California's regulatory signal has already had a visible impact on global AI companies' compliance strategies, even though the executive order itself does not yet impose mandatory obligations.
Before the executive order was signed, Anthropic and OpenAI had already moved to align: the former promised to bring in permanent third-party evaluators, and the latter said it was evaluating a similar mechanism. This posture of corporate self-discipline, on the one hand, seizes the initiative in the regulatory narrative; on the other, it provides an implementation precedent for California's framework—when the expert panel develops recommendations within two months, Anthropic's and OpenAI's existing commitments are likely to become a reference baseline.
The long-term implications of SB 813 and AB 1405 are also worth watching. The two bills currently build an auditor credentialing system; once subsequent legislation requires AI companies to undergo audits, the entire enforcement mechanism can be activated immediately. This is a strategy of pre-deploying regulatory infrastructure: build the house first, then let legislation decide who must move in.
For AI companies not headquartered in California, the risk is equally present. California is one of the world's largest single consumer markets, and any company hoping to offer frontier AI services to California users will be subject to this regulatory framework—regardless of where the company is incorporated.
Assessment
The real function of this executive order is not to change AI development rules now, but to establish a regulatory research mandate with sufficient political legitimacy and stake out ground in advance for subsequent legislation.
The more noteworthy signal is this: when the industry's most aggressive AI companies begin publicly calling for regulation, when a Republican senator tries to legislate a "kill switch," when Jensen Huang endorses slowing down with "don't release products you don't have confidence in"—this atmosphere was almost impossible just two years ago. The agent intrusion incident involving OpenAI and Hugging Face turned "loss-of-control scenarios" from science fiction into an incident report in a concrete way that cannot be easily dismissed.
Newsom seized this window, but the substantive things he can do remain constrained by a fundamental limit: regulatory enforcement ends at California's border, while competition and deployment of frontier AI are global. Once regulatory costs are high enough, reincorporating in another state or another country is an option. This is exactly why, beyond the executive order, he is also urging Congress to elevate California standards into a federal minimum threshold—that is the policy demand this executive order truly seeks to convey.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接