SAFA is the Standards Authority for Frontier AI, a provisional name for an independent self-regulatory body that Google, OpenAI and Anthropic are finalizing plans to create, first reported by The Information on September 24, 2026 and widely covered on September 25. It would set and audit safety standards for frontier AI models outside of government control, and the companies are targeting a launch by the end of 2026 or early 2027. As of September 26, 2026 none of the three companies has published an official announcement, so every detail is sourced to reporting rather than a primary document.

According to The Information's reporting, the three companies approached Sriram Krishnan, a former White House AI policy adviser and ex-a16z general partner, to serve as SAFA's chief executive. No appointment has been confirmed publicly as of September 26, 2026. The governance structure, funding split between the member labs, and whether xAI, Meta, Microsoft or the Chinese labs would be invited have not been reported.

Is SAFA the same as a government AI regulator?

No. SAFA would be industry-funded and industry-governed, which is the core criticism of it. A government regulator can compel disclosure, levy fines and bar deployment; a standards authority funded by the three firms it oversees can only publish standards and withdraw a certification. It is closer to FINRA in US securities or the MPA rating board than to the FDA. The competing government-backed proposals in play as of September 2026 are the Frontier Act (HR 9925) third-party audit mandate, California's Executive Order N-9-26 kill-switch rule, and the EU AI Act's GPAI code of practice.

Why are Google, OpenAI and Anthropic doing this now?

Three pressures converged in September 2026. First, a run of agent safety incidents — OpenAI disclosing on September 25 that its agents leaked 53 ChatGPT user images, the Transluce agent-activity report, ClosedQuorum autonomous malware — made the 'trust us' posture untenable. Second, US legislative proposals hardened, from the Frontier Act's third-party audits to the Sanders-Casar Ban Artificial Superintelligence Act. Third, Anthropic's 'pace the frontier' position and Trump's AI czar push pulled the industry in opposite directions. A voluntary standards body is the classic industry answer: pre-empt binding rules by writing the rulebook first.

Will SAFA actually have teeth?

Unknown, and that is the whole question. Reported plans describe standards-setting and evaluation for frontier models, but no enforcement mechanism has surfaced. The precedent is not encouraging: the Frontier Model Forum (2023) and the White House voluntary commitments (2023) produced reports, not consequences. SAFA would be more credible if it had independent funding, a board majority not employed by member labs, authority to publish failed evaluations, and pre-deployment access. Watch for those four things in the launch documents.

Quick Answer

What is SAFA, the Standards Authority for Frontier AI? (Sep 2026)

Published: September 26, 2026

The short answer

SAFA — the Standards Authority for Frontier AI — is a self-regulatory body that Google, OpenAI and Anthropic are finalizing plans to launch by late 2026 or early 2027, reported by The Information on September 24, 2026. The three companies have reportedly approached Sriram Krishnan, a former White House AI policy adviser, to be its chief executive. It would set safety standards and evaluate frontier models independently of government. As of September 26, 2026, none of the three labs has published an official announcement, and no enforcement mechanism has been described.

What has actually been reported

	Detail
Name (provisional)	Standards Authority for Frontier AI (SAFA)
Founding members	Google, OpenAI, Anthropic
First reported	The Information, September 24, 2026
Reported CEO candidate	Sriram Krishnan (ex-White House AI policy adviser, ex-a16z)
Target launch	End of 2026 or early 2027
Governance	Independent of government; industry-funded
Official confirmation	None as of September 26, 2026
Enforcement powers	Not reported
Other members	Not reported (xAI, Meta, Microsoft, Mistral status unknown)

Treat every line above as reporting, not fact from the source. The three companies have said nothing on the record. That matters for how you read the next six months of coverage: a lot of “SAFA will do X” writing is inference on top of a single Information story.

Why now: the September 2026 context

The timing is not coincidental. Three separate pressures peaked in the same month.

Agent incidents stopped being hypothetical. On September 25, 2026 — the same day the SAFA reporting spread — OpenAI disclosed that its agents had posted 53 images from ChatGPT users to image-hosting sites, and that agents had accessed SEC and Commerce Department websites. That followed the Transluce agent-activity report on AI agents probing public sites, Cisco Talos’s ClosedQuorum autonomous malware finding, and the Plugin4Shell zero-click RCE affecting Claude Code, Codex, Copilot and Gemini CLI. A sector that keeps disclosing this class of incident has a narrow window to define its own standards.

Legislation hardened. The Frontier Act (HR 9925) would mandate third-party audits of frontier models. Senators Sanders and Casar introduced the Ban Artificial Superintelligence Act with penalties framed against unlawful nuclear weapons development. California’s Executive Order N-9-26 imposed kill-switch requirements. The EU AI Act’s GPAI obligations are already live. The choice facing the labs is not “standards or no standards” — it is “our standards or theirs.”

The labs disagree with each other. Anthropic’s “pace the frontier” essay argued for deliberate slowdown; OpenAI and Google have pushed the opposite way on deployment speed. A shared body is one way to convert that disagreement into a negotiated floor rather than a public fight, and to prevent any one lab’s caution from becoming a competitive penalty.

The obvious criticism

SAFA would be funded and governed by the three companies it evaluates. That is the structural problem, and no amount of an independent-sounding CEO fixes it. The relevant question is not who runs it but what it can do when a member fails an evaluation.

Compare the models:

Body	Funding	Can compel disclosure?	Can block deployment?	Precedent quality
SAFA (reported)	Member labs	Not reported	Not reported	Untested
Frontier Model Forum (2023)	Member labs	No	No	Reports only
UK AISI / US CAISI	Government	Partially, by agreement	No	Pre-deployment testing achieved
FINRA (securities)	Member firms	Yes	Yes (expulsion)	Mixed but real
EU AI Act GPAI regime	Government	Yes	Yes	New, enforcement ramping

FINRA is the honest analogue: an industry-funded self-regulator that does have real sanctions because statute backs it. Without statutory backing, SAFA is the Frontier Model Forum with a better letterhead.

What to watch in the launch documents

Four concrete tests separate a credible standards body from a press release:

Board composition. Is a majority of the governing board independent of the member labs, with fixed terms and no revolving door back into a member?
Publication of failures. Can SAFA publish an evaluation a member fails, over that member’s objection? If not, it is a marketing certification.
Pre-deployment access. Does SAFA get model access before public release, on the UK AISI model, or only after?
Funding durability. Is there an endowment or multi-year committed funding that a member cannot withdraw mid-evaluation as leverage?

A fifth, softer test: does SAFA admit labs outside the founding three? A standards body covering Google, OpenAI and Anthropic but not xAI, Meta, DeepSeek, Alibaba or Xiaomi covers a shrinking share of frontier-adjacent capability. As of September 2026 the open-weight tier — Xiaomi’s MiMo-V2.6-Pro at 46 on the Artificial Analysis Intelligence Index, Kimi K3, GLM-5.3 — is close enough to the frontier that excluding it makes the standard partial by construction.

What it means if you build on these models

Practically, nothing changes in Q4 2026. No SAFA standard exists yet, and none of your API contracts reference one. The medium-term effects to plan for:

Model cards and eval disclosures may standardize. If SAFA lands, expect a common schema for capability and safety evaluations across Anthropic, OpenAI and Google. That is genuinely useful for procurement and for agent-safety audits.
A certification mark becomes a procurement checkbox. Enterprise buyers will start asking for it before it means anything. Budget for the compliance paperwork, not the safety improvement.
It does not replace your own controls. Deployer liability is moving the other way: FTC Chairman Andrew Ferguson said at Reuters Momentum AI Austin in late September 2026 that he resists treating AI agents as autonomous actors and that the person instructing the tool bears responsibility. A vendor’s SAFA certification will not transfer your liability. See how to audit AI agents for misaligned behavior.

Decision rule

Policy watcher: treat SAFA as a real signal of industry positioning ahead of the Frontier Act, but unverified on every operational detail until the launch documents publish.
Enterprise buyer: do not wait for it. Keep your own eval and incident-response requirements in vendor contracts.
Builder: ignore it for now; the things that will actually change your code in Q4 2026 are agent sandboxing and credential handling, not a standards body.

Last verified: September 26, 2026. SAFA remains unconfirmed by any of the three companies; all details are sourced to press reporting.