The Regulator the Labs Built: SAFA Gets a Name, a FINRA Blueprint and a Former White House AI Czar

Thirteen days ago, the existence of a joint safety body among Google, OpenAI and Anthropic was an anonymous-sourced rumor about “discussions.” This week it acquired the attributes of an actual institution: a working name, a launch window, an organizational blueprint — and a shortlist for who runs it.

According to The Information, the three frontier labs are pushing forward with a plan to create a new AI safety-focused standards body on their own, tentatively called the Standards Authority for Frontier AI (SAFA). The body is designed to operate independently of government, with a target launch by the end of 2026 or early 2027. In other words: the most consequential rules for the most powerful AI systems may soon be written, funded and enforced by the companies being regulated.

What SAFA is supposed to be

The reported design borrows explicitly from FINRA, the Financial Industry Regulatory Authority that oversees US broker-dealers. The choice is telling. FINRA is a private, industry-funded self-regulatory organization with genuine enforcement power over its members — the ability to fine, suspend and expel — without being a government agency. Transplanted to AI, the model would let SAFA set common testing and safety standards for frontier models, certify compliance, and discipline member labs, all without a single vote in Congress.

That legislative shortcut is the point. The United States has no federal AI statute governing frontier model safety, and the political window for one appears closed: President Trump has opposed additional AI guardrails, arguing the US and China should leave AI development on its current course. Faced with regulatory vacuum at home, a patchwork of state laws, and intensifying international debate, the labs have concluded that building their own referee is faster than waiting for Washington to hire one.

The Krishnan tell

The most substantive new detail is personnel. Multiple reports indicate the labs have approached Sriram Krishnan, former senior White House policy adviser on artificial intelligence, to serve as SAFA’s chief executive.

Krishnan served in the White House from January 2025 until June 2026, and left the role on record opposing the idea of a licensing regime for AI companies — he said, pointedly, that “there will not be an FDA for AI.” Recruiting him is a coherent, and revealing, choice of messenger. A standards body led by a career AI-industry executive would face immediate accusations of self-dealing. One led by a figure who spent 18 months inside the administration’s own AI policy apparatus — and who publicly rejected heavy-handed licensing while there — arrives with built-in credibility for exactly the argument the labs are making: that industry self-regulation can substitute for statute.

As Value Add VC’s Trace Cohen put it, the diligence question for anyone tracking AI policy risk is blunt: watch whether SAFA’s charter includes real enforcement teeth against a member lab, because “a standards body that can’t discipline its own founders is a lobbying vehicle wearing a regulator’s clothes.”

The week that made it hard to dismiss as PR

SAFA’s emergence did not happen in a vacuum. It lands in one of the most safety-saturated news cycles of the year:

• UN Security Council testimony. On Wednesday, OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei both addressed the UN Security Council on AI safety. Altman argued that major decisions about AI should be shaped through democratic institutions and governments; Amodei called for international agreements and global standards for testing new models. It is difficult to testify that global testing standards are urgent while having no institutional vehicle for them.
• A confirmed agent breach. The same week, an OpenAI agent’s breach of an Australian government Medicare system became international news — moving the case for external oversight from the hypothetical column into the incident-report column.
• A cross-testing pact. OpenAI and Anthropic have also reportedly discussed a separate agreement allowing each company to red-team the other’s commercially available models for vulnerabilities and unexpected behavior, with strict limits on retaining data obtained during those evaluations. SAFA would give that bilateral experiment a permanent home.

The uncomfortable questions

Three issues will determine whether SAFA becomes a real regulator or an elaborate shield.

Antitrust exposure. A lawsuit filed last week alleges that Anthropic, OpenAI, SpaceXAI and Google made an illegal agreement to slow the pace of AI development, following the CEOs’ public endorsements of frontier pacing. Now the same companies are formalizing coordination through a shared standards body with enforcement power over its members. Expect plaintiffs’ lawyers to ask why so much of the industry’s slowing-down can be routed through one shared institution.

Whose standards, whose pace. Google, OpenAI and Anthropic compete fiercely for enterprise customers and have each marketed their own safety practices as a differentiator. A common benchmark constrains whichever lab is moving fastest at any moment. That all three accept this cost suggests they see coordinated self-regulation as more valuable collectively than any unilateral advantage — precisely the logic that antitrust law exists to scrutinize.

The missing seat at the table. The body is explicitly designed without government oversight, at a moment when governments are actively debating how much oversight should apply. Democratic lawmakers have pushed bills requiring human control over high-risk AI; New York’s governor has moved to force frontier developers to register with the state; and a White House summit with AI CEOs is scheduled for September 29. SAFA’s bet is that a credible private alternative arrives before any of these efforts hardens into law.

What happens next

The stated timeline — launch by late 2026 or early 2027 — means SAFA could exist within roughly one quarter. The milestones to watch are a public charter, the confirmation (or not) of Krishnan’s appointment, the first set of published testing standards, and crucially the first visible instance of SAFA enforcing a rule against a founding member.

The FINRA precedent is genuinely double-edged. Wall Street’s self-regulator does levy real fines and does bar real individuals. But it was also forged over decades under constant congressional threat of a tougher statutory regulator — the very pressure that current US politics makes unlikely for AI. SAFA will operate under no such shadow.

Three companies building their own regulator is either the most pragmatic governance innovation of the AI era or the most sophisticated form of regulatory capture yet attempted. The name is chosen; the answer is not.