Who actually regulates frontier AI in 2026?

No single body. As of September 2026 four regimes overlap: the EU AI Act, which is binding law with GPAI obligations already in force; California's Executive Order N-9-26, which imposes kill-switch requirements on models deployed in the state; the Frontier Act (HR 9925), a US bill proposing mandatory third-party audits that is not yet law; and SAFA, an industry-funded standards body Google, OpenAI and Anthropic were reported on September 24-25, 2026 to be forming, targeting launch by late 2026 or early 2027. Only the first two can compel anything today.

Is SAFA a replacement for government AI regulation?

No, and that is the criticism of it. SAFA would be funded and governed by the frontier labs it evaluates, with no statutory authority to compel disclosure, levy penalties or bar deployment. Its closest working analogue is FINRA in US securities — an industry self-regulator that has real sanctions only because statute backs it. Without that backing, SAFA's realistic output is standards documents and a certification mark, which is the Frontier Model Forum model with better branding. Its practical function is to shape what binding rules look like before they arrive.

What does the Frontier Act (HR 9925) require?

The Frontier Act is a US bill that would mandate third-party audits of frontier AI models, moving evaluation from voluntary self-reporting to independent assessment. It is the most direct competitor to SAFA's positioning: if enacted, an industry-funded standards body becomes at best a compliance vendor rather than the governing layer. As of September 2026 it is a bill, not law, and its passage is the single biggest variable in whether SAFA matters.

Which regime applies to me if I build on AI models rather than train them?

The EU AI Act, most likely, and sooner than you expect. Its obligations attach to deployers as well as providers, scaled by risk category, and they apply to anyone putting a system on the EU market regardless of where the company sits. California's EO N-9-26 kill-switch rule targets model deployment in state. SAFA and the Frontier Act are aimed at frontier model developers, so as a builder you would encounter them indirectly through vendor attestations. Separately, US FTC enforcement is already live on deployer responsibility without any new statute.

What did the FTC say about AI agent liability in September 2026?

FTC Chairman Andrew Ferguson, speaking at the Reuters Momentum AI Austin conference in late September 2026, said he resists describing AI agents as autonomous actors with wills and desires of their own, and that developers and those instructing AI tools should bear liability. He used a hammer analogy: if someone hits a person with a hammer, the law examines the person, not the tool. He also argued existing product liability and consumer protection law should be tested before new AI legislation, and cautioned against adopting European-style rules. That is an enforcement posture available today, requiring no new statute.

Quick Answer

SAFA vs Frontier Act vs EU AI Act vs California EO: Who Regulates AI?

Published: September 26, 2026

The short answer

Four regimes are competing to govern frontier AI as of September 2026, and only two of them can compel anything. The EU AI Act is binding law with GPAI obligations in force. California’s Executive Order N-9-26 imposes kill-switch requirements on models deployed in state. The Frontier Act (HR 9925) would mandate third-party audits but is still a bill. SAFA — the Standards Authority for Frontier AI reported on September 24-25, 2026 — would be industry-funded, industry-governed, and has no described enforcement power. Meanwhile the FTC is applying existing consumer protection law to deployers without waiting for any of them.

Side by side

	SAFA	Frontier Act (HR 9925)	EU AI Act	California EO N-9-26
Type	Industry self-regulation	US federal bill	Binding EU law	State executive order
Status (Sep 2026)	Reported, unconfirmed	Not enacted	In force	In force
Backers	Google, OpenAI, Anthropic	Congress	EU institutions	California governor
Funded by	Member labs	Taxpayer	Taxpayer	Taxpayer
Core mechanism	Standards + evaluation	Third-party audits	Risk tiers + GPAI obligations	Kill-switch requirement
Can compel disclosure	Not reported	Yes, if enacted	Yes	Yes, in scope
Can block deployment	Not reported	Yes, if enacted	Yes	Yes, in state
Penalties	None described	Statutory, if enacted	Up to % of global turnover	State enforcement
Applies to deployers	Unclear	Frontier developers	Yes	Model deployment
Geographic reach	Member labs	US	Extraterritorial (EU market)	California
Target launch	Late 2026 / early 2027	Unknown	Phased, live	Live

The honest ranking by teeth

1. EU AI Act. It is law, it is extraterritorial, and the penalties are a percentage of global turnover. If you put a system on the EU market you are in scope regardless of where you incorporate. It also reaches deployers, not just model providers, which is why it is the regime most likely to affect a company that merely builds on APIs.

2. California EO N-9-26. An executive order is weaker than statute and faster to issue. The kill-switch requirement is narrow but concrete, and California’s market position means compliance tends to become the national default. See what the California AI kill switch executive order requires.

3. FTC enforcement under existing law. Not a new regime at all, which is the point. Chairman Ferguson’s late-September 2026 position — that developers and instructing parties bear liability, and that existing product liability and consumer protection law should be tested before new AI statutes — means enforcement can start tomorrow under authorities that already exist. For most companies this is the nearest-term legal exposure, and it lands on deployers.

4. Frontier Act (HR 9925). Real teeth if enacted — mandatory third-party audits would restructure the evaluation market — but it is a bill. See the Frontier Act explained.

5. SAFA. Last, because as of September 26, 2026 it does not exist, has not been confirmed by any of the three companies, and no enforcement mechanism has been described. See what SAFA is.

What SAFA is actually for

Reading SAFA as a safety mechanism misreads it. Read as a positioning move it makes immediate sense.

The industry faces the Frontier Act’s audit mandate, the Sanders-Casar Ban Artificial Superintelligence Act with penalties framed against unlawful nuclear weapons development, California’s kill-switch order, and EU obligations already biting. A voluntary standards body launched before binding rules arrive does three useful things for its members: it produces a technical vocabulary that later legislation tends to adopt, it gives legislators an off-ramp (“industry is handling it”), and it sets a floor so that one lab’s caution — Anthropic’s “pace the frontier” position, for instance — does not become a unilateral competitive penalty.

That is not cynical, exactly. Industry standards bodies have produced real safety outcomes in aviation and finance. But they did it with statutory backing and independent funding, and SAFA as reported has neither.

The four tests

Whether SAFA becomes FINRA or becomes the Frontier Model Forum depends on four things visible in its launch documents:

Board independence — is a majority not employed by member labs, with fixed terms?
Failure publication — can it publish an evaluation a member fails, over that member’s objection?
Pre-deployment access — does it see models before release, on the UK AISI model?
Funding durability — endowment or committed multi-year money a member cannot withdraw as leverage?

A fifth: membership breadth. A body covering Google, OpenAI and Anthropic but not xAI, Meta, Mistral, DeepSeek, Alibaba or Xiaomi covers a shrinking slice of frontier-adjacent capability. The open-weight tier reached 46 on the Artificial Analysis Intelligence Index in September 2026 with Xiaomi’s MiMo-V2.6-Pro, and an Apache-2.0 agentic model that runs on one consumer GPU shipped the same week. Standards that only bind three American labs do not bind the capability frontier.

What to do, by role

If you train frontier models: all four matter, in the order above. Engage with SAFA’s formation if invited, because the standards written now become the statutory language later.

If you build products on AI APIs: the EU AI Act deployer obligations and FTC enforcement are your real exposure. SAFA and the Frontier Act reach you only as vendor attestations. Do not let a future SAFA certification substitute for your own controls — Ferguson’s hammer analogy cuts directly at the “our vendor was certified” defense.

If you procure AI: write your own eval and incident-disclosure requirements into contracts now. Every regime above is either unsettled or under-enforced, and contract terms are the only instrument you control. Ask specifically about agent egress boundaries and credential handling, which is where September 2026’s incidents concentrated — see how to give AI agents credentials without leaking them.

If you are forecasting policy: watch the Frontier Act’s committee progress. Its fate determines whether SAFA is the governing layer or a compliance vendor underneath a statutory one.

Decision rule

Binding today: EU AI Act, California EO N-9-26, existing FTC authority.
Binding if enacted: Frontier Act.
Not binding, possibly influential: SAFA.
Applies to you as a builder rather than a lab: EU AI Act deployer duties and FTC deployer liability. Those two, first, before anything in the news cycle.

Last verified: September 26, 2026. SAFA remains unconfirmed by Google, OpenAI or Anthropic; all details are sourced to press reporting of The Information’s September 24, 2026 story.