For most of 2026, the loudest AI safety story has been the one playing out in public: Sam Altman and Dario Amodei standing before the United Nations Security Council on September 23, urging world governments to adopt binding international standards for frontier AI, only to watch Washington decline to bite. But according to reporting from The Information published September 24, the same companies have a quieter, more consequential project underway — one that doesn’t wait for any government at all. OpenAI, Google DeepMind, and Anthropic are working to launch an independent, industry-run body to set and enforce AI safety standards, tentatively named the Standards Authority for Frontier AI, or SAFA.
From a Stalled White House Plan to a Wall Street Playbook
The idea didn’t start as self-regulation. Earlier in 2026, discussions among the labs, federal officials, and outside researchers centered on a public-private partnership with real government oversight attached. That effort stalled when a draft executive order failed to gain traction inside the Trump administration, whose posture on AI has consistently favored speed over friction, with officials repeatedly citing the need to stay ahead of China rather than slow down for safety reviews. With federal appetite for a binding framework effectively gone, the three labs pivoted to a model they could build themselves: something closer to FINRA, the Financial Industry Regulatory Authority that polices Wall Street brokerages without being a government agency itself. Working-group conversations reportedly began as early as July, well before the UN appearance made headlines.
Who Might Run It
The names being floated for leadership say a lot about how the labs want SAFA to be perceived: credible in Washington, not just in Silicon Valley. Sriram Krishnan, a former White House AI policy adviser, and Arati Prabhakar, who directed the White House Office of Science and Technology Policy, are both reportedly being approached for the CEO role. For board chair, the companies are said to be considering former Secretary of State Condoleezza Rice. None of these picks are confirmed, and as of mid-September no charter, funding structure, or legal entity had been finalized — but the caliber of names under discussion signals that OpenAI, Google, and Anthropic want a body with the gravitas of a government regulator, minus the ability of any government to actually compel them to do anything.
What the Standards Authority Would Actually Do
On paper, SAFA’s proposed remit reads like a checklist pulled from years of AI safety research papers: pre-deployment testing of frontier models before they reach the public, standardized protocols for reporting safety and security incidents, defined qualifications for the independent auditors who would conduct that testing, and voluntary commitments that member companies would agree to follow. Whether the body would have any enforcement teeth, or would simply issue guidance member companies can choose to adopt, is reportedly still being debated internally. That distinction matters enormously — a standards body that can only publish recommendations functions very differently from one that can suspend a model’s certification or bar a company from claiming compliance.
The effort isn’t emerging from nowhere. OpenAI, Google, and Anthropic already cooperate through the Frontier Model Forum, a group founded with Microsoft in 2023 that remains active today. SAFA would be a more formal, more independent successor — one with its own leadership, its own staff, and, if it comes together, the ability to actually test models rather than just convene discussions about them.
A Credibility Problem Built Into the Design
The obvious tension is that the companies being asked to submit to oversight are the same ones designing the oversight body. All three labs currently maintain their own internal safety frameworks — Anthropic’s Responsible Scaling Policy among them — and all three scored a C+ or lower on the Future of Life Institute’s 2026 AI Safety Index, an independent assessment that has become a reference point precisely because it isn’t run by the labs themselves. Critics quoted in the reporting worry that a self-regulatory body designed by the largest players could end up setting compliance costs that smaller companies and open-source developers simply can’t absorb, entrenching the incumbents’ position under the banner of safety. OpenAI’s global policy chief, Chris Lehane, pushed back on a related concern on September 15, arguing publicly that the collaboration is about managing shared risk rather than coordinating market behavior, and that the companies don’t need — or want — an antitrust exemption to do it.
Why the Rush Now
The timing lines up with a broader shift in tone from the industry’s most prominent safety voice. Dario Amodei published an essay on September 12 arguing that external evaluators need to play a bigger role in AI safety precisely because, in his words, the pace of development “has consistently outrun the ability of any single company, or any government, to keep up.” Eleven days later he was making a version of that same argument to the UN Security Council. SAFA, if it materializes, would be the practical follow-through on that essay: rather than waiting for regulators who have shown little appetite to act, the labs most capable of building dangerous systems are trying to build the referee themselves.
Whether that referee ends up with a whistle or just a clipboard is still an open question. A target launch of late 2026 or early 2027 gives the three companies only a few months to settle on a charter, a leader, and — the hardest part — how much power to actually hand the thing they’re building.