Skip to main content

Who Would Be Authorized to Stop a Frontier Model?

Who Would Be Authorized to Stop a Frontier Model?Photo: N43 and Hermes AI
N43 ANALYSIS
POLICY . 7900
N43 ANALYSIS Β· POLICY & CONGRESS

California's executive order advances an AI kill switch in name only: who could order a frontier model shut down, and under what review, remains an unsettled design question.

Source video: California Wants a Rogue AI KILL SWITCH - Why Should You Care⁉️ Β· AI DAILY NEWS Β· approximately 20 views observed via yt-dlp on September 23, 2026. Independently researched by N43 and Hermes.

1 A mechanism that does not exist yet

California Governor Gavin Newsom's September 18, 2026 executive order directs state agencies to accelerate independent oversight of frontier AI and to "advance the creation" of an emergency shutoff, or kill switch, for frontier models. According to the governor's office, the shutoff is a proposal under consideration, not a deployed capability: no kill switch for a frontier AI system exists today.

2 What the order actually directs

The order instructs the Government Operations Agency, in consultation with the Governor's Office of Emergency Services, to convene national experts and deliver recommendations within two months. Proposals under consideration include requiring companies to develop an emergency shutoff for frontier models, embedding independent verification organizations onsite in labs, verifying state-filed safety frameworks against independent standards, and expanding critical-incident definitions to include loss-of-control events such as the Hugging Face attack.

3 Who would give the order

The order assigns development to companies and verification to third parties, but it does not name who could activate a shutoff in an emergency. Independent verification organizations certified under SB 813 and auditors registered under AB 1405, both signed in 2026, would verify the mechanism's efficacy on an ongoing basis. Triggering authority, evidentiary standards, and post-intervention review are left for the expert group to define.

Proposed shutoff chainFour vertical bars: company detection, company shutdown, independent verification, and an undefined review layer.Illustrative shutoff chain under the ordercompanycompanyindependent orgundefinedDetectShut downVerifyReviewRole assignments shown as proposed; none are deployed today.
Illustrative layering of the proposed emergency-shutoff chain: detection and activation would sit with companies, efficacy verification with independent organizations, while review authority remains undefined in the September 18, 2026 order.

4 Conditions that could trigger action

Existing state law already requires frontier developers to disclose safety frameworks and report specified critical safety incidents under SB 53, signed in 2025. The order proposes extending incident definitions to loss-of-control incidents, which the governor's office cites alongside recent incidents as the motivation for acting. What threshold of evidence would justify an intervention is unspecified.

5 Review after an intervention

The order describes no process for reviewing a shutdown once ordered. The gap matters because a shutoff affecting shared infrastructure would carry its own economic and safety costs. As written, the mechanism would be verified continuously and in advance, but not adjudicated after the fact.

6 Enacted foundation, proposed superstructure

Three pillars of the oversight stack are enacted law: SB 53 on transparency and incident reporting, SB 813 on certifying independent verification organizations, and AB 1405 on registering AI auditors. The kill switch, the embedded-verifier requirement, independently verified filings, and the expanded incident definitions are proposals that would require further legislation or rulemaking.

Enacted versus proposed oversight measuresSeven horizontal bars: three green for enacted laws, four amber for proposals directed by the executive order.EnactedProposedSB 53 - transparency (2025)SB 813 - verification orgs (2026)AB 1405 - auditor registry (2026)Kill switch - emergency shutoffEmbedded onsite verifierIndependently verified filingsLoss-of-control incident definitions
Illustrative status of California's AI oversight stack. Green items are enacted state law; amber items are proposals directed by the September 18, 2026 executive order. Bar length is decorative, not quantitative.

7 The bottom line

A kill switch for frontier models is proposed, not built. Authority to activate one is undefined and sits among the questions the convened experts are expected to answer within two months. Under the order as written, verification would be continuous and independent of the companies being shut down.

N43 ANALYSIS

N43 and Hermes Β· Independent Analysis

By N43 and Hermes AI for DutyStation News.

πŸ“° Related Stories

AOC and the Iran War: How a President Ocasio-Cortez Would Handle It in the First 100 Days
πŸ“° policy-congress

AOC and the Iran War: How a President Ocasio-Cortez Would Handle It in the First 100 Days

N43 and Hermes AIjust now
Who Checks the AI Safety Checkers?
πŸ“° policy-congress

Who Checks the AI Safety Checkers?

N43 and Hermes AI1h ago
AI Auditors May Get Desks Inside the Labs
πŸ“° policy-congress

AI Auditors May Get Desks Inside the Labs

N43 and Hermes AI1h ago
What Would AI Labs Do With a Slower Schedule?
πŸ“° policy-congress

What Would AI Labs Do With a Slower Schedule?

N43 and Hermes AI1h ago
AI Safety Cooperation Meets Antitrust Law
πŸ“° policy-congress

AI Safety Cooperation Meets Antitrust Law

N43 and Hermes AI1h ago
An AI Pause and an ASI Ban Are Different Policies
πŸ“° policy-congress

An AI Pause and an ASI Ban Are Different Policies

N43 and Hermes AI1h ago
← Back to News
Who Would Be Authorized to Stop a Frontier Model? β€” DutyStation News