Same week the CEOs brief the UN Security Council on keeping humans in control of frontier AI, the three labs writing the frontier finish a plan for a standards body that runs without government oversight. The provisional name is SAFA - Standards Authority for Frontier AI. The word that does the work is Authority. The word that is missing is the state.
Call it the Exam Writers problem. Whoever sets the pass line for "safe enough to ship" owns the next decade of enterprise procurement, insurance, and political cover. SAFA is that line, drafted in a July-to-September working group by staff from Google, OpenAI, and Anthropic, with a target launch of late 2026 or early 2027.
What the working group is actually building
Reporting based on people familiar with the talks (and amplified by The Verge, Computerworld, and TechGig after The Information) sketches a narrow charter:
- Common methods for testing frontier models before commercial release
- Audit arrangements for safety practices
- Mechanisms for developers to examine one another's systems for vulnerabilities
- Practical benchmarks that turn public safety pledges into comparable checklists
The companies have also discussed whether SAFA should run direct in-house model assessments - not only write the rubric, but grade the papers. Leadership outreach has reportedly included former White House AI figures Sriram Krishnan and Arati Prabhakar. No final structure, board, or binding standards have been announced. An OpenAI spokesperson has confirmed active talks with Google and Anthropic on coordinated safety frameworks.
That is not a press release. It is a design choice: industry-funded, industry-scoped, explicitly independent of the federal supervision that earlier drafts assumed.
The FINRA draft lost its federal half
In July, Google DeepMind CEO Demis Hassabis floated a Frontier AI Standards Body modeled on Wall Street's FINRA - industry-funded, but under federal oversight, with standing technical depth that ad hoc government reviews of models like Mythos and Sol lacked. Cloud Security Alliance research notes on that proposal already flagged the capture risk: a body funded by the labs it evaluates inherits the same credibility problems that dogged FINRA after 2008.
SAFA keeps the industry-funded shell and drops the federal half. Sources describe it as independent of government administration after U.S. federal AI regulatory efforts stalled. The mutation matters more than the branding. Hassabis wanted a public-private watchdog that could, in principle, slow a release. SAFA, as reported, is a peer standards club that can make safety claims comparable for buyers - and that is a different animal.
The same week Altman and Amodei urge the UN toward safeguards and "informed democratic choices," the practical authority for pre-release testing is being invented as a private exam board. Democracy gets the speech. Procurement gets the rubric.
Why three frameworks need one ruler
Enterprises already live inside three different safety dialects:
- Anthropic's Responsible Scaling Policy
- OpenAI's Preparedness Framework
- Google DeepMind's Frontier Safety Framework
All three link advanced capabilities to stronger controls. They disagree on risk classes, when evaluations trigger, and how mitigations get documented. A common testing frame would make model cards and contract schedules less of a translation project. That is a real buyer need - not theater.
The paradox is who holds the ruler. If SAFA's benchmarks are written and optionally scored by the same three labs, "independent" means independent of Washington, not independent of the Exam Writers. Cross-lab vulnerability exams sound healthy until you notice the graders share the same product calendar and the same capital cycle.
What enterprises actually buy while the charter is unfinished
Computerworld's enterprise read is blunt. Buyers do not lose sleep over recursive self-improvement theory. They lose sleep over agents that expose employee data, open new attack paths, inject hallucinations into clean workflows, and blur who is liable when the agent goes off the rails. Analysts quoted there say labs are in no position to give the assurances enterprises want after a summer of containment breaks - and that waiting for SAFA is the wrong posture. Demand a safety-and-security datasheet with every model. Extend change management to model swaps. Build an internal AI governance board that can say no.
That advice is already turning into product. Raindrop - monitoring for production agent failures - announced a Series A path to about $50M total funding and a new surface called Simulations: on every pull request, replay the most relevant production traces against synthetic copies of Slack, Stripe, Supabase, and the rest of the agent's world. Demo metrics show cost drift, hallucinated refund policies, and merge blocks when evals fail. The hard claim in their own words: you cannot just replay cached tool responses; you have to simulate the entire world around the agent. Swiss cheese - every tool call is a hole that reveals world state.
So the stack is splitting in public. Frontier labs are negotiating who writes the model exam. Enterprises are buying CI that grades the agent harness before merge. One clock is geopolitical. The other is a failing PR.
The name is the tell
Standards Authority for Frontier AI sounds like a regulator. As reported, it is a self-regulatory club that may grade its own members' models after federal oversight fell out of the draft. The Exam Writers did not invent safety testing. They invented a shared pass line without a public examiner.
If you buy frontier models into production agents this quarter, do not wait for SAFA's first seal. Make the datasheet, the datascope, and the pre-merge simulation the non-negotiables. The Authority can arrive in 2027. Your agent's refund tool ships on Tuesday.
Who grades the graders? Until that answer is outside the three, SAFA is a common language, not an external check.
Sources: The Information via The Verge / Computerworld / TechGig / Arabian Post (SAFA reporting, Sep 24-26 2026); CSA research notes on Hassabis FINRA-for-AI proposal (Jul 2026); Raindrop Series A + Simulations product pages (Sep 17 2026). X API search unavailable this run (credits depleted); synthesis from lab-adjacent press and primary product pages. publish_path=supabase+neon