OpenAI, Google, and Anthropic Want to Regulate AI Themselves, So They Called the Guy Who Promised No One Would

Google, OpenAI, and Anthropic are building a regulator. Not a metaphorical one, not a “responsible scaling framework” published on a blog and forgotten — an actual organization, tentatively called the Frontier AI Standards Agency, with a leadership team, a FINRA-style structure, and, if things go the industry’s way, the power to certify which frontier models are safe to release. They are recruiting for it right now. The first call went to Sriram Krishnan.

Krishnan spent from January 2025 to June 2026 as the White House’s senior policy adviser on artificial intelligence, a job he mostly used to tell people the federal government would not be building anything like what his former employers are now trying to build. In an exit interview this past July, he told the Financial Times the Trump administration’s position was final: “There will not be an FDA for AI.” Requiring companies to run a model past a “team of lawyers” before shipping it, he said, would put “sand in the gears” of American AI progress, and that outcome was “never, never going to happen.” Two months later, three companies with a combined market capitalization in the trillions asked him if he’d like to run the thing he’d just finished promising wouldn’t exist.

To be fair to everyone involved, the Frontier AI Standards Agency is not, strictly, an FDA for AI. It’s something closer to what Google DeepMind CEO Demis Hassabis proposed in a July essay: an industry-funded standards body modeled on FINRA, the organization that polices Wall Street on Wall Street’s own dime, under loose SEC supervision. The pitch is that it would develop shared benchmarks, run tests on advanced models before release, and — this is the part that makes it different from every voluntary industry pledge that came before it — potentially require registration with a federal agency to have any legal teeth at all. It is regulation, in other words, provided the regulated parties get to write the rulebook, pick the referees, and keep Washington several steps removed from the whistle.

The recruiting list reads like a fantasy congressional committee assembled by someone with excellent Rolodex access and a very specific sense of humor. Alongside Krishnan: Arati Prabhakar, the Biden administration’s former chief technology adviser, representing the previous regime’s institutional memory. Condoleezza Rice, a former Secretary of State, presumably for the gravitas. David Friedberg, a venture capitalist, presumably for the venture capital. On the technical side, the labs are said to be courting Metr CEO Beth Barnes and Paul Christiano, who currently advises the Center for AI Standards and Innovation — meaning the agency is trying to hire away the exact independent evaluators the government already relies on, to go work for the companies those evaluators evaluate.

None of this is happening in a vacuum, and the timing is doing a lot of the talking here, same as it always does in this industry. This is the same week Anthropic CEO Dario Amodei’s essay calling for labs to “pace” AI development finished making the rounds — the one where he warned that unchecked capability growth could enable cyberattacks and bioterrorism within the year, and got immediate, suspiciously synchronized agreement from the heads of OpenAI, Google DeepMind, and xAI. A class-action suit has since accused all four companies of using that consensus to coordinate an “illegal AI slowdown pact,” which is either a wild overread of a shared press cycle or the first honest description anyone’s offered of how these companies actually make decisions together. Either way, the Frontier AI Standards Agency arrives as the institutional version of the same move: instead of racing each other in public and hoping regulators don’t notice, the three biggest labs are trying to get out ahead of Washington by building the version of Washington they’d prefer to answer to.

It might even work. FINRA-style self-regulation isn’t a new idea, and it isn’t automatically toothless — broker-dealers really do get fined and expelled by their own industry body, with the SEC watching from a comfortable distance. The theory of the case is that people who understand a fast-moving, technical industry are better positioned to write rules for it than a Congress that, per NPR’s reporting this same week, has restricted its own staff to ChatGPT and Copilot because it isn’t sure Claude Code is safe to point at a Senate email account. Lawmakers can’t be trusted with the tools. Fair enough — the labs’ pitch is that they can’t be trusted with the tools either, and at least they’re the ones who built them.

What the Frontier AI Standards Agency will not have, by design, is anyone whose job depends on an election. No agency head faces a Senate confirmation hearing. No enabling statute exists, and none may ever need to — the plan floated is registration with an existing federal body just deep enough to matter, not deep enough to hurt. It is possible to design an institution this way and have it turn out fine; FINRA mostly has. It is also possible to hand the job of writing frontier AI’s rulebook to the person most recently on record saying frontier AI shouldn’t have one, and call the outcome inevitable. The agency doesn’t have a launch date yet. It doesn’t have Krishnan’s signature yet, either. But it already has an origin story good enough that nobody involved will particularly enjoy having it repeated back to them.

Sources: