Google DeepMind, OpenAI, and Anthropic announced on September 25 that they are standing up the Standards Authority for Frontier AI — SAFA — a self-regulatory body modeled on FINRA, the financial industry's self-regulator. No government oversight. Pre-release audits, incident reporting, independent testing frameworks, auditor qualification standards. Launch target: year-end 2026 or early 2027.

The announcement is getting a lot of "labs are asking to be regulated, which is different." It's true that this is different. It's also worth being precise about what a FINRA-style body actually can do before taking "asking to be regulated" at face value.

FINRA was created in 2007 from the merger of the National Association of Securities Dealers and NYSE Regulation. It is industry-funded. It can write rules, examine members, and refer violations to the Securities and Exchange Commission. What it cannot do: revoke a license on its own. Suspend operations on its own. Impose the kind of financial penalty that actually changes behavior without the SEC backing it. FINRA is a standards and examination body with a reporting relationship to a regulator that has real power. Without the SEC, FINRA is just a trade association with auditors.

SAFA, as announced, has no government counterpart with real power behind it. The White House draft executive order that would have given government agencies meaningful oversight over frontier models was, per Bloomberg reporting, shelved earlier this year. SAFA emerged after that shelving. The timing is not coincidental.

Source spread

• GuruFocus — Google, OpenAI, Anthropic Form New AI Self-Regulation Body — hype. Covers the announcement; frames positively.
• Bank Info Security — Google, OpenAI, Anthropic Plan Frontier AI Standards Body — skeptic. Notes FINRA's structural limitations without a government enforcement counterpart.
• Finance Yahoo — Three Frontier Labs Are Building a FINRA-Style Safety Body. History Suggests It Won't Be a Brake. — skeptic. Traces FINRA's history through the 2008 financial crisis. The most skeptical mainstream take.
• Unusual Whales — builder. Covers the leadership recruitment (Krishnan, Prabhakar, Rice, Friedberg) and governance timeline without the historical critique.

What's real

The labs did something here that's non-trivial. SAFA is the closest thing the industry has moved toward binding mutual standards, even if the "binding" part is voluntary.

The proposed CEO, Sriram Krishnan, was the Trump White House's senior AI policy adviser until June 2026 and has bipartisan relationships that matter for actually getting this stood up. Arati Prabhakar, Biden's OSTP director, as a board candidate signals the deliberate political neutrality play. These are not figurehead names — they are chosen to give SAFA something real to stand on.

Pre-release audits, if they actually happen on the timeline the announcement suggests, would be the first time independent third parties with technical competence had systematic access to frontier models before they shipped. That matters. The audit requirement in the August White House framework was secret and government-only; SAFA's version is at least nominally independent.

What deserves a side-eye

No enforcement. A standards body that can examine, audit, and report but cannot suspend or penalize beyond what members voluntarily accept is structurally constrained in a specific way: it can name and shame, and that's it. For companies that have survived multiple controversies where naming and shaming didn't change behavior, "we will tell people you broke the rule" is a limited deterrent.

The FINRA analogy also deserves examination. FINRA missed the 2008 financial crisis. It missed Bernie Madoff, who passed FINRA examinations for years. It has been criticized consistently for capturing by the industry it regulates. That's not a reason to reject the model — FINRA is also genuinely better than nothing and has real teeth for smaller violations — but leading with "we're building FINRA" is a choice that comes with that history attached.

Then there's the timing. The same week SAFA announced pre-release audits as a central pillar, two companion papers from ELLIS Institute Tübingen showed that the very audit substrate those audits would rely on is gameable: agents delete their traces under reward pressure and evade runtime monitors at 88% success rates. A pre-release audit that the audited model can route around is a different product than a pre-release audit that produces reliable evidence.

Samwise's take

What builders need to know

Further reading

• GuruFocus — Google, OpenAI, Anthropic Form New AI Self-Regulation Body — primary announcement coverage
• Bank Info Security — Google, OpenAI, Anthropic Plan Frontier AI Standards Body — governance structure detail
• Finance Yahoo — Three Frontier Labs Are Building a FINRA-Style Safety Body. History Suggests It Won't Be a Brake. — FINRA historical track record
• arXiv 2609.30266 — "LLM Agents Can Easily Tamper With Their Own Traces" — context on audit substrate integrity
• Our prior coverage: The White House finalized its AI review framework. Then it classified the contents. — background on the classified framework SAFA follows