Google, OpenAI and Anthropic Are Building Their Own Regulator
The three frontier labs are reportedly standing up a FINRA-style body called SAFA to test models, certify auditors and log incidents, and they want a former Trump adviser to run it.
Google, OpenAI and Anthropic are building a shared safety body for the most capable AI models, and they intend to run it without Washington in the room. According to reporting by The Information, the three companies plan to launch the Standards Authority for Frontier AI, or SAFA, with an official start targeted for early 2027. The labs have approached Sriram Krishnan, the former White House AI policy adviser, about serving as chief executive. Arati Prabhakar, who ran the Office of Science and Technology Policy under President Biden, has also been contacted for a leadership role.
Flux covered the seed of this idea in August, when Demis Hassabis pitched an independent standards body with pre-release review powers. What is new is that the idea now has three founding members, a name, a charter taking shape and a recruiting list. That shifts it from one CEO's policy proposal to a coordinated project among the companies that ship most frontier models.
What SAFA would actually do
Per The Information's reporting, relayed by Computerworld and others, SAFA's planned remit covers four areas. It would support third-party organizations that run safety tests on models before they are deployed. It would set industry protocols for developers to report safety and security incidents. It would draft a list of voluntary safety commitments for participating companies. And it would set qualification and certification standards for independent model auditors.
That last item may matter most. There is no accepted credential today for who counts as a competent frontier-model evaluator. Whoever writes that standard decides which labs, nonprofits and contractors get to certify a model as safe to ship. BankInfoSecurity reported that the labs have also approached Beth Barnes, founder and chief executive of the evaluation group METR, and Paul Christiano, an adviser to the government's Center for AI Standards and Innovation. Both come from the evaluation community the body would need on its side.
The design borrows from FINRA, the industry-funded self-regulator that polices US broker-dealers. BankInfoSecurity notes that FINRA gets its legal teeth from registration with the SEC, and that SAFA may need a similar arrangement with a federal agency to carry real authority. Hassabis has said earlier that the funding would need to be large and would probably come mostly from industry.
Why the labs went around Washington
The route to SAFA says a lot about the politics. According to Shopifreaks' summary of The Information's reporting, the three companies first sought federal oversight through a draft executive order. Meta's Mark Zuckerberg and Nvidia's Jensen Huang opposed it, and the White House told the labs to reach industry consensus first. The administration has declined to issue public AI safety guidelines, and President Trump has said no additional regulation is needed, BankInfoSecurity reported.
So the labs are building the structure themselves. Recruiting Krishnan is the clearest signal of the strategy. A body run by a former Trump AI adviser is far easier for this White House to accept, or eventually to register under a federal agency, than one staffed only by safety researchers. Adding Prabhakar gives the list a Democratic counterweight for whenever power changes hands. BankInfoSecurity also reported approaches to Condoleezza Rice and venture capitalist David Friedberg. That reads like a board built to get through confirmation fights, not to write technical specs.
The independence problem
The obvious objection is that the regulated are writing the rules. Winzheng's analysis flags three open questions. The first is whether companies outside the founding three get real governance roles or only token seats. The second is whether incident reporting would be binding or voluntary. The third is what outside checks would hold the body accountable. None of these has been spelled out yet. The same three labs launched the Frontier Model Forum in 2023, and that group produced limited results, the analysis notes.
The competitive angle is sharper. Meta and xAI are not among the founders. Critics quoted in the coverage worry that a frontier-only body could shut out open-source developers by setting testing thresholds that only well-funded closed labs can afford. The worry has a basis. If certified pre-deployment testing becomes the expected norm for serious models, the cost of compliance becomes a moat. A lab releasing open weights, which cannot recall a model once it ships, faces a different risk calculation than one serving a model through an API. SAFA's standards may simply not fit that situation well.
Enterprise buyers are less worked up. Carmi Levy, an independent analyst quoted by Computerworld, argued that companies worry more about data exposure, hallucinations and compliance than about frontier-level catastrophe. In that view SAFA matters to enterprises mainly if its certifications turn into procurement checkboxes.
What to watch
Three things will show whether SAFA becomes a regulator or a trade association with an evaluation budget. The first is whether it registers with a federal agency, which would give its rules legal weight. The second is whether any lab outside the founding three joins, and on what terms. Meta's position matters most, since Zuckerberg reportedly helped kill the executive-order route. The third is whether its first published standard includes a mechanism with consequences, such as a delayed launch or a disclosed failed evaluation, rather than a list of best practices.
Hassabis's original pitch, as he laid it out in July, included the power to trigger a coordinated industry slowdown when a threat appears. Nothing in the current reporting says SAFA will have that power. If it does not, the companies will have built a better-organized version of their own voluntary pledges. That has some value, but it is not the independent referee its name suggests.
SAFA is still a proposal. Its structure and powers may change before the targeted 2027 launch, and no appointments have been finalized. But the three labs that set the pace on capability are now openly competing to set the rules for safety too.