SAFA: The Labs' FINRA-Style Self-Regulator Takes Shape — Without Government, Meta, xAI or Nvidia
Reported Sept 26, 2026 (TheStreet, via Yahoo Finance). The cross-lab safety standards body has a name — the Standards Authority for Frontier AI (SAFA) — and a target launch window of late 2026 or early 2027. The framing sentence of the story is the whole story:
"Three of the most powerful companies building artificial intelligence are working on something unusual. Rules for themselves, written by themselves. The federal government is largely on the sidelines."
What SAFA would do
Per the report, key guidelines include: third-party testing before models ship, rules for how labs report safety incidents, and a framework that turns voluntary safety pledges into something that can actually be checked. The idea traces to a July proposal from Demis Hassabis modeling it on FINRA (Wall Street's self-regulator), industry-funded.
Telling origin detail: the three labs originally wanted a public-private partnership with federal oversight, but that "stalled after a draft White House executive order was put on hold. The administration told the companies to reach industry consensus first" — which is how you get a regulator born from a request not to regulate. OpenAI's Chris Lehane confirmed the trilateral talks in mid-September; Altman publicly backed the effort.
Who runs it, and who's not in it
- CEO shortlist: Sriram Krishnan (current administration AI adviser) and Arati Prabhakar (former Biden science office) — "the bipartisan casting is not accidental."
- Chair shortlist: Condoleezza Rice and venture capitalist David Friedberg.
- Out: "Meta, xAI, and Nvidia pushed back publicly at Dreamforce on Sept. 15. SAFA is a deal among three labs, not a consensus across the whole field."
The enforcement gap
"FINRA can actually punish brokers. SAFA, without government registration, probably cannot. It would set standards but may not necessarily enforce them."
The story's closing question is the right one: "Can a body funded by the companies it polices ever actually police them? Or does it just give everyone cover to keep doing what they were already doing?"
Where this fits
SAFA runs parallel to, and independently of, the White House voluntary accord signed September 29 (Trump Hosts AI CEOs; Six Frontier Firms Sign Voluntary "Constitution" — Then Renames AI "Super Intelligence" and Launches America.gov) — two self-regulatory tracks, one White House-endorsed, one deliberately outside government. It is also distinct from the reported OpenAI–Anthropic bilateral stress-testing talks (still unconfirmed; watch remains open). Whether SAFA ever adopts the kind of published, comparable metrics Anthropic now releases (see Anthropic's Automation Trajectory Goes to the IMF: Claude Projected to "Lead" 80–90% of Frontier Model Training by Next Year) is the test of whether it's a transparency regime or a shield. The Information corroborates the story's thrust ("Google, OpenAI and Anthropic AI Safety Group Takes Shape") but the full piece is paywalled.