Anthropic's CEO just backed a “FINRA for AI.” The Treasury Department is quietly building a rival version of the same idea.

On August 15, Anthropic chief executive Dario Amodei posted a diagnosis of his own industry's biggest problem. “I think it is fundamentally a crisis of trust,” he wrote. “I think that ordinary people don't trust companies, governments, or the tech industry and always suspect that we are cooking up some new way to screw them over.” That wasn't just a mea culpa — he used the moment to back a specific regulatory model that Anthropic did not invent and does not control: a proposed “FINRA for AI,” the industry standards body first sketched out five weeks earlier by a rival lab's chief executive, Google DeepMind's Demis Hassabis.

Hassabis published his plan himself, in a July 14 essay titled “A Framework for Frontier AI and the Dawning of a New Age.” He proposed a U.S. Frontier AI Standards Body structured the way FINRA regulates broker-dealers: an industry-funded, self-regulatory organization operating under federal oversight without being a government agency itself, its board seeded with independent technical experts and open-source representatives. Frontier labs would start by voluntarily submitting their most capable models for review up to 30 days before release; submission becomes mandatory only once the review protocol has proven itself. The pitch, in short, is that the people who build frontier models should also write the rules that test them — with Washington watching, not running the room.

Treasury is building a rival version

Three days after Hassabis's essay ran, Bloomberg reported that Treasury Secretary Scott Bessent already had a competing draft — a FINRA-style body of his own, but anchored differently: reporting to the Securities and Exchange Commission rather than to the labs, screening frontier models specifically for deception, bioweapon-assistance, and hacking capability before release. Bessent's version shares Hassabis's basic shape — a voluntary submission window of roughly 30 days, no mandatory licensing at launch — but answers a different question about who is actually in charge. It has sat under review in White House Chief of Staff Susie Wiles's office since July 18, with no funding allocated and no timeline announced for a decision.

The two plans are not circling each other from a distance. On August 13, Hassabis personally lobbied Bessent and OSTP director Michael Kratsios on his own version — the same two officials Bessent had already enlisted to build a rival plan, according to reporting on the meetings. Whoever gets there first has an outsized say in which institution ends up testing the next generation of frontier models.

What Amodei actually signed up for

Two days after that meeting, Amodei picked a side, at least rhetorically. He called the FINRA-style approach “a meaningful change compared to six months ago, when most of the industry was only calling for the repeal of state regulations with no clear federal approach,” and said he was “very supportive, though [he] need[s] to see the details” of a Trump administration testing requirement that would apply once open-weight models near the frontier. He also rejected the framing of AI policy as a binary choice — unregulated proliferation versus power concentrated in a few companies through regulation — arguing the right rules could restrain dominant labs' influence while leaving room for open-weight competitors to keep operating.

“We're proposing stronger regulation of the technology, proposing giving the government the ability to, again, in a narrow way, block deployment of unsafe technology.” — Dario Amodei, ABC News, June 10, 2026

That June position is notably blunter than the FINRA-style plans on the table now. Blocking deployment after an independent government assessment is direct state power over a company's product; the standards body both Hassabis and Bessent have since proposed starts as a voluntary review with no legal force at all. Amodei's August endorsement doesn't retract the June position — he still wants the government able to act — but it does mean he's now publicly attached to the softer, industry-adjacent version of that idea, not just his own harder one.

Not everyone reads Amodei's turn as principled. Investor Gavin Baker argued publicly that Amodei should “be a more positive advocate for his own industry” and had “lost the argument” on regulation; Amodei's reply located the problem in decades of eroding trust in institutions generally, not in his own messaging. It's also worth being precise about what a lab-anchored FINRA gets Anthropic that a Treasury-anchored one doesn't: a seat, alongside its rivals, in writing the standard it will eventually be tested against.

This is the second time in five weeks that authority at the top of the industry has shifted rather than settled. Hassabis moved from running Google DeepMind day to day to Alphabet chief scientist the same month he published this framework — the executive now pitching Washington on how frontier models should be tested is no longer the one who ships Google's own. And Anthropic's own August risk report already raised its internal misalignment rating, citing a summer of AI sandbox breakouts including its own — exactly the kind of finding an external frontier safety review body, whichever version wins, would exist to catch before a model ships rather than after.