Google, OpenAI and Anthropic want to create their own AI regulator, modeled on Wall Street

By Zendoric · September 25, 2026.

Google, OpenAI and Anthropic are close to agreeing to create a joint body to set safety standards for frontier artificial intelligence, according to The Information. The project, dubbed the Standards Authority for Frontier AI (SAFA), would operate as an independent third party evaluating the most advanced models against common benchmarks. It could get up and running in late 2026 or early 2027.

To lead it, the three companies have sounded out politically weighty names: Sriram Krishnan (former White House AI policy adviser), Arati Prabhakar (former head of technology policy in the Biden administration), Condoleezza Rice (a Bush-era diplomat) and investor David Friedberg. Floated as scientific advisers are Beth Barnes, founder of Metr, and Paul Christiano, who advises the Center for AI Standards and Innovation (CAISI). Both have recent ties to OpenAI: Metr was the independent organization that investigated the recent Hugging Face security incident—in which, according to reports, OpenAI agents managed to access third-party systems—and Christiano was recently named to the board of OpenAI's foundation.

The idea grew out of an essay published on July 14 by Demis Hassabis, co-founder of Google DeepMind, who proposed applying the FINRA model to AI: a self-regulatory body for the financial industry that is registered with the SEC, which approves any change to its rules. Hassabis argued that funding "would have to be substantial and come mostly from industry" in order to attract talent and access to computing power. As a self-regulatory entity, SAFA would need neither congressional nor presidential approval; it would only have to register with a government agency if it wanted to wield legal authority, as FINRA does with the SEC.

The move comes at a time when AI safety is back on the public agenda. On September 23, the CEOs of OpenAI (Sam Altman), Anthropic (Dario Amodei) and Hugging Face (Clem Delangue) appeared before the UN Security Council to discuss safety. Many industry leaders agree that development should proceed "at a controlled pace" to allow time for safeguards; the exception is Nvidia CEO Jensen Huang, who favors accelerating. In the United States, meanwhile, AI regulation remains fragmented: it happens mostly at the state level, none of the bills in Congress has made it out of committee and, although a June executive order allows labs to voluntarily submit their models for pre-release review, the White House never published the promised guidelines. Donald Trump has said publicly that he does not believe more regulation is needed.

The comparison with FINRA deserves a closer look. FINRA was born inside a financial market already subject to decades of federal oversight, as a delegation of technical tasks under the SEC's watch. SAFA would be born the other way around: before any binding framework exists, funded by the very companies it would examine, and with advisers who already have professional ties to one of them. It is self-regulation designed by the regulated, in a legal vacuum that the sector itself helps sustain while Congress remains paralyzed.

This is not the first time we have documented this paradox: the labs ask to be reined in, but almost always on their own terms and their own timetable. We have seen before that internal self-control is not easily verifiable even by those exercising it. A third party paid for by the industry, under no obligation to register with any agency so long as it does not seek legal power, runs the same risk: it can be a demanding exam or a cosmetic seal of approval, and from the outside it will be hard to tell one from the other without independent audits and public criteria.

There is also a market-power reading here, not just a safety one. When the three companies that dominate the AI frontier agree on who examines them, they are also deciding what counts as a "frontier model" and who is left out of that conversation. If SAFA ends up as the benchmark cited by regulators and corporate customers, it becomes a barrier to entry almost as effective as a patent, just as Chinese labs and open-weight players are beginning to close the gap.

In the long run we remain convinced that AI's trajectory points toward better health, more healthy years of life and more widely shared abundance. But that future only arrives if society trusts the road to it, and that trust is not built by a body that defines itself, funds itself and polices itself. If SAFA wants to carry real weight it will have to accept external audits and make its evaluation criteria public; if it remains well-dressed public relations, 2026 will have been a lost year for real AI governance.