SAFA Would Let Frontier Labs Write the Auditor Scorecard Procurement Uses
Google, OpenAI, and Anthropic are forming the Standards Authority for Frontier AI to qualify auditors and testing practices, which could redefine what independently evaluated means in enterprise AI contracts.
What changed
Reporting in late September 2026 described Google, OpenAI, and Anthropic advancing an industry led body tentatively named the Standards Authority for Frontier AI, abbreviated SAFA. People familiar with the talks told The Information the group could launch by late 2026 or early 2027 with a remit to support third party pre deployment testing, define incident reporting practices, and set qualification requirements for independent auditors.
The effort sits alongside, not instead of, the 29 September White House accord in which six tech leaders pledged layered internal and external controls. SAFA would operationalize what counts as a credible external review when labs self police frontier models.
Why it matters
Enterprise procurement teams already ask vendors for safety evaluations. Without a shared auditor roster, independently evaluated becomes marketing noise. SAFA, if it ships with teeth, would tell buyers which testing shops and benchmark suites count when assessing models such as Gemini 4 Argon, Claude Mythos, or OpenAI Astra tiers.
The risk is structural: the labs designing the standard compete for the same customers. Critics note Meta, xAI, Microsoft, Mistral, and open weight builders may sit outside the initial frontier definition, which could turn SAFA into a competitive moat dressed as governance. Buyers should demand published charter text, funding independence, and conflict rules before treating SAFA certification as equivalent to regulator oversight.
Who is affected
Chief information security and AI governance officers must rewrite vendor questionnaires to ask for SAFA aligned audit artifacts once charter details appear.
Independent evaluators such as METR affiliates face either partnership or displacement depending on SAFA auditor qualification rules.
Policy staff tracking the White House accord should read SAFA as the industry's attempt to pre empt mandatory federal testing regimes.
What to do next
Pause automatic renewal clauses that reference generic independent AI safety evaluations. Instead require vendors to specify auditor identity, test suite version, and raw report availability pending SAFA charter publication.
What to watch
A joint public announcement from Google, OpenAI, and Anthropic, published SAFA charter and board appointments, and whether U.S. Commerce CAISI coordinates or competes with SAFA testing mandates.
Sources
- Primary. The Information via Dealroom summary, Google, OpenAI and Anthropic AI Safety Group Takes Shape (24 September 2026). SAFA naming, launch timing, and auditor qualification scope attributed to people familiar.
- Secondary. The Verge, OpenAI, Anthropic, and Google planning AI safety organization (24 September 2026). Corroborating report on SAFA remit and 2027 launch target.