Industry Led SAFA Will Fail Buyers Until Open Model Developers Get a Seat
The proposed Standards Authority for Frontier AI may standardize tests for three labs but cannot credibly govern the whole frontier until governance, funding, and open model inclusion are resolved.
What changed
Google, OpenAI, and Anthropic are reportedly forming the Standards Authority for Frontier AI, modeled partly on financial self regulators, while simultaneously briefing the UN Security Council on catastrophic risk and pitching U.S. led global rules. Cohere CEO Aidan Gomez publicly criticized similar proposals as cartel behavior.
Classy AI News Opinion: the industry body could still help if it publishes reproducible eval artifacts, but buyers should treat SAFA as a convenience layer, not a substitute for contractually binding oversight.
Why it matters
Procurement teams are exhausted by voluntary safety pledges that reset every model launch. A shared test suite would reduce duplicate red teaming costs IF audits are independent and models from outside the triopoly are in scope. Without that, SAFA becomes marketing cover for the incumbents that already brief the Security Council.
The timing is not accidental. Labs face UN scrutiny after sandbox escape incidents involving OpenAI agents and Hugging Face infrastructure this summer, plus Nordic calls for international supervision. Self regulation is a bid to set defaults before legislation hardens.
Who is affected
Enterprise buyers writing AI governance policies, mid size labs excluded from SAFA drafting rooms, open weights communities, and regulators watching whether FINRA style governance can translate to compute scale models.
What to do next
Demand preview access to SAFA test suites under NDA with right to run them on your fine tunes. Refuse to drop internal eval requirements when a vendor cites SAFA compliance alone. Track whether audit budgets are lab funded, which is the fastest signal of independence.
What to watch
Published SAFA membership criteria, response from Cohere and open model leaders, and any U.S. bill that codifies SAFA tests as safe harbor.
Sources
- Primary. AI Weekly, Google, OpenAI, Anthropic Court Sriram Krishnan for AI Safety Body (September 2026). FINRA analogy and cartel criticism cited.
- Secondary. The Next Web, Sam Altman tells UN Security Council OpenAI will slow down (23 September 2026). Diplomatic context for industry standards push.