OpenAI, Google and Anthropic Want a FINRA for AI. Without Statutory Backing, It's a Rulebook With No Referee.

The three largest AI labs are working on a body called the Standards Authority for Frontier AI, or SAFA: voluntary, industry-funded, meant to set and verify safety standards for frontier models. OpenAI’s Chief Global Affairs Officer Chris Lehane confirmed in mid-September that the company had spent “several weeks” in talks with Anthropic and Google DeepMind, per reporting from TheStreet and TechRepublic. The target launch window is late 2026 or early 2027.

We build agents on these labs’ models, so who writes the rules they ship under is a planning input for us. Here is what exists, and what doesn’t.

What exists: a plan

As of now there is no published charter, no confirmed leadership and no standards. Names are circulating: Sriram Krishnan, an AI adviser to the current administration, and Arati Prabhakar, who ran the Biden-era White House science office, are reportedly shortlisted for CEO. Condoleezza Rice and David Friedberg are reportedly in the running to chair. Treat all of that as rumor until a charter appears.

The idea traces to a July 2026 proposal from Google DeepMind CEO Demis Hassabis, who suggested modeling the body on FINRA, the industry-funded broker regulator on Wall Street. The labs originally wanted a public-private partnership with federal backing. That stalled after a draft White House executive order was shelved, with the administration reportedly telling the industry to reach consensus among themselves first. SAFA is the fallback.

The planned functions are third-party testing before models ship, standardized rules for reporting safety incidents, and a framework meant to make today’s voluntary pledges, like the labs’ responsible-scaling policies, verifiable instead of self-certified.

The structural problem

FINRA can discipline and fine brokers because it has statutory backing. SAFA, without government registration, almost certainly cannot. It may be able to set standards but not enforce them in any binding way.

That is the same shape as the White House’s one-page voluntary accord from September 29: commitments to audits and shared standards, with enforcement left to the companies that signed. SAFA would add an institution on top. It would not obviously add a penalty.

The questions critics are raising are specific:

• Can a body funded by the three companies it oversees hold them accountable? Reporting indicates the group is still debating whether evaluations will be done in-house or by independent third parties. That choice will decide whether the exercise has any credibility.
• Is this an industry consensus or a deal among three? Meta, xAI and Nvidia publicly pushed back at Dreamforce on September 15. That makes SAFA the three biggest labs agreeing on safety, with everyone else outside the room.
• Is it a moat? If SAFA sets high third-party testing bars, compliance costs could be prohibitive for smaller players and open-source developers. The entities writing the rules would also be the ones judged against them. That is the sharpest version of the regulatory-capture critique, and it matters most if enterprise customers or insurers start treating SAFA compliance as a baseline expectation.

The case that it’s more than PR

There are signs of real urgency inside at least one of the three labs. Anthropic CEO Dario Amodei published an essay on September 12, “We Must Pace the Frontier,” arguing AI capability is advancing faster than the industry’s ability to control it. Sam Altman has publicly backed the effort as something the industry needs to do together.

There is also precedent that cuts both ways. OpenAI, Google, Microsoft and Anthropic set up the Frontier Model Forum in 2023, per TechCrunch. SAFA is a more ambitious, FINRA-styled version of the same impulse, which raises the obvious question of why the earlier effort wasn’t enough.

What this means if you ship agents

Nothing changes today, because nothing exists yet. But two scenarios are worth planning for:

1. SAFA gets teeth or becomes a de facto gate. Model access, pricing or API terms could eventually reflect SAFA-style testing requirements, and smaller model providers could be squeezed. If you depend on a non-Big-Three or open-weight model, that is a supplier-risk question.
2. SAFA stays reputational. Then the practical control layer remains yours: egress rules, scoped credentials and your own audit trail around agents, regardless of what a vendor certifies.

We don’t think the second scenario requires waiting to see how the first resolves.

What we’re watching

• Whether SAFA ends up with any enforcement mechanism, or is purely voluntary.
• Whether Meta, xAI or others join, or it hardens into a closed Big Three club.
• Who is named CEO and chair, and whether that signals a safety regulator or an industry shield.
• Whether the shelved federal executive order resurfaces once SAFA either works or visibly fails to self-police.