Late on September 24, The Information, citing people familiar with the matter, reported that Google, OpenAI, and Anthropic are moving forward with a new AI safety standards organization, tentatively named the Standards Authority for Frontier AI (SAFA). The three companies want it up and running by the end of this year or early next year, without government oversight. Coverage from GuruFocus and CGTN adds that, in the absence of a new government regulatory framework, the three chose industry self-regulation to fill the gap, and that SAFA plans concrete testing and auditing systems rather than general ethical guidelines.

[1][2]

The nut graf: what matters here is not another pledge of responsible AI from the big three, but that self-regulation has moved from declaration to institutional design — the reported SAFA comes with executable details: pre-deployment third-party safety testing, incident reporting rules, and independent auditing requirements. That distinguishes it from the Frontier Model Forum founded in 2023, a declarative industry body that included Microsoft. The new organization is being advanced by Google, OpenAI, and Anthropic alone, Microsoft absent, and explicitly without government endorsement. The Information's report says the three companies initially sought a public-private partnership with federal oversight, but that path stalled over lack of consensus within the AI industry and shifting government priorities — so they built their own. 54 55 The reported design splits into three layers. On scope, SAFA focuses on safety standards for advanced AI models: not general ethical principles but pre-deployment third-party safety tests and incident reporting rules — a signal that safety is becoming an enforceable workflow rather than a statement. On governance, it plans to operate independently to fill the regulatory vacuum, and that is precisely its soft spot: a self-regulatory body without statutory power must build legitimacy through industry acceptance. GuruFocus's report notes it will need to coordinate with existing regulators and win broader industry support before it can be recognized as a legitimate standard-setting authority. On membership, three companies lead and Microsoft is not on the list, contrasting with the Frontier Model Forum's four-member lineup (Microsoft included) and giving the body from day one a flavor of core players writing the rules. 56 57 Put the story in its timeline: over the past week OpenAI and Anthropic shipped new models on the same day (the GPT-6 line and Claude Opus 5.5), with at least eight models appearing in seven days (a daily AI roundup opened with exactly that), while safety standards remain at the stage of voluntary statements. The three companies pushing SAFA now is effectively an admission that waiting for government rulemaking is not working, and an attempt to claim the power to define standards through self-regulation. The tension worth watching: the standard-setters and the regulated are the same companies. Independence of testing and auditing, transparency of incident reports — all of that must eventually be delivered under external scrutiny. The reports so far describe only the institutional skeleton, no governance details, which is exactly what to track next.

[1][2]
A late-night corner of an open-plan office: in the foreground a long table with a stack of white papers and three dark tablets (pure graphics, no text), a pen resting on the papers; midground, a meeting room behind a glass partition with empty chairs around a round table and a blank whiteboard; background, city lights outside floor-to-ceiling windows and a row of dark monitors. No people, no text.
Three companies writing their own safety standards, who supervises the supervisors, AI-generated illustration, not a news photo