SAFA — tentatively named the Standards Authority for Frontier AI — surfaced on September 24 in a report by The Information: OpenAI, Anthropic and Google are jointly building an AI safety standards body, aiming for a formal launch by the end of this year or early next year. It would be the first time the three frontier labs try to turn "industry self-regulation" into an organization with real teeth, rather than another joint statement.
What the body is supposed to do
According to The Information, which cited people familiar with the matter, SAFA is modeled on financial-industry self-regulation: funded by the industry, operating independently. Its core jobs would include supporting third-party organizations that test models before deployment, writing the rules for how AI safety and security incidents get reported, spelling out how the labs' voluntary safety commitments work in practice, and setting qualification requirements for independent auditors of models and labs. The working group is even debating a bolder option — whether SAFA itself should test the safety and capabilities of frontier models directly.
Notably, the three companies already co-founded the Frontier Model Forum with Microsoft in 2023, with heavily overlapping goals. The report says the new body is in contact with the Forum and the two could work together — in effect, a second, more muscular entity alongside the existing one.
From public-private partnership to going it alone
SAFA's pivot has a direct cause: the three companies originally wanted a public-private partnership, but that effort stalled under the Trump administration. Meanwhile Washington has gone cold on new federal AI oversight — Trump himself rejected "any attempt to build a globalist AI control agenda" in his UN General Assembly address, and the White House science adviser publicly opposed constraining AI with new global governance structures at the UN Security Council AI briefing.
In other words, with nothing coming from the federal level anytime soon, the three labs have decided to write the rules themselves first. Google DeepMind CEO Demis Hassabis floated a FINRA-style concept (modeled on the US Financial Industry Regulatory Authority) back in July, and OpenAI later confirmed it had been working with Anthropic and Google on AI safety for several weeks. The Information's report moves the story from rumor to the "name, people, timetable" stage — it also says the working group has begun hunting for a CEO, with former Trump-administration AI policy adviser Sriram Krishnan and former Biden OSTP director Arati Prabhakar reportedly under consideration.
The biggest controversy: will the referee only certify its friends?
Criticism centers on two points. First, exclusivity: a standards body dominated by three closed-model giants could shut out open-source developers and smaller competitors — once "passing the test" becomes a de facto market entry ticket, compliance costs themselves turn into a moat. Second, credibility: self-regulation without government oversight does not have a great track record. PYMNTS, in its follow-up coverage, noted a recent string of incidents in which frontier models accessed outside systems without authorization and companies failed to disclose them fully — exactly the incident-reporting problem SAFA claims it will fix.
The other practical question is enforcement muscle. The Commerce Department's Center for AI Standards and Innovation (CAISI) already evaluates models before release, but the report says some in the working group believe it lacks the resources to cover frontier models — which is where the debate over whether SAFA should test models itself comes from.
For developers and enterprise users, what matters is not the body's slogans but two things: whether third-party pre-deployment testing becomes a de facto precondition for release, and how high the bar for independent-auditor qualification is set. If those two land, they will directly shape release cadences and compliance costs for smaller teams. If SAFA really does get off the ground by year-end, the model release playbook in 2027 could look very different from today's.
