[UPDATE] The AI Labs' Self-Regulator Takes Shape

When the largest AI companies signaled this month that they wanted to police their own work, the plan was still a sketch. It now has a name, a model to copy, and a shortlist of candidates to run it. Google, OpenAI and Anthropic are working to stand up an independent body called the Standards Authority for Frontier AI, or SAFA, that would test advanced models and set shared safety benchmarks before they reach the public.

We reported earlier in September that the three labs were circling an industry standards group. The reporting since then, led by BankInfoSecurity, fills in what it might actually look like.

Built to look like Wall Street's self-cop

SAFA would run as a self-regulatory organization modeled on the Financial Industry Regulatory Authority, the private body that oversees brokers and securities firms in the United States. That structure matters. Like FINRA, SAFA could operate without an act of Congress or a presidential order, though it may still need to register its legal authority with a government agency.

Its job, as described, would be to develop common standards for judging frontier models, support outside groups doing pre-deployment testing, and set out how the voluntary safety promises the labs already make should work in practice. Funding would be substantial and would come mostly from the industry itself, which is the part critics keep circling back to.

Who might run it

The labs have approached two people with government pedigree. One is Sriram Krishnan, who advised the second Trump White House on AI until June 2026 and left after rejecting the idea of a licensing regime, saying plainly that there would be no "FDA for AI." The other is Arati Prabhakar, a senior technology official in the Biden administration. No chief executive has been appointed, and the group could launch late this year or in early 2027.

The obvious tension

A referee paid for by the teams it officiates invites the obvious question. The same week SAFA came into focus, the heads of these companies were warning the UN Security Council that poorly managed AI could threaten the world, and state attorneys general were pressing Congress to write binding rules rather than trust the industry to grade itself. Anthropic's Dario Amodei has spent months arguing for a slower, more careful pace, a position the White House has publicly rejected.

SAFA is a bet that the labs can set a credible bar before anyone in government sets one for them. Whether a body they fund can say no to the models they build is the test that counts, and it has not been run yet.