Voluntary Safety Working Groups Still Leave Enterprise Buyers Without Binding Tests
OpenAI confirmed weeks of safety coordination with Anthropic and Google DeepMind, but enterprises still lack enforceable procurement criteria while working groups remain private.
What changed
On 15 September 2026, OpenAI global affairs chief Chris Lehane told reporters the company had been working with Anthropic and Google DeepMind for several weeks on AI safety coordination. OpenAI said it does not believe an antitrust waiver is required for rivals to cooperate on safety matters. Coverage from Bloomberg, CNBC, and SiliconANGLE tied the disclosure to Dario Amodei's 12 September essay urging frontier labs to pace development and to working group meetings reported since July 2026 on a voluntary standards body modeled partly on FINRA.
Separately, Google DeepMind's Demis Hassabis published a September essay through the DeepMind Institute elaborating a Standards Body proposal with pre release testing and possible coordinated slowdowns among frontier labs. As of mid September, no joint public framework document from all three labs had shipped.
Why it matters
Enterprise buyers face two parallel tracks: vendors publicly aligning on safety pacing, and procurement teams still signing model contracts without shared pass fail tests. Voluntary working groups may reduce tail risk for labs, but they do not automatically produce auditable criteria you can embed in an RFP. Until benchmarks, evaluator independence, and disclosure rules are published, safety talk remains reputational rather than contractual.
The antitrust framing also matters for procurement lawyers. Lehane's no waiver needed stance is a lab view, not legal clearance for customers to demand synchronized slowdown clauses across vendors.
Who is affected
Chief information security officers, procurement leads buying frontier APIs, insurance and risk teams modeling agent deployments, and regulators watching whether self regulation precedes mandatory rules.
What to do next
Draft agent procurement requirements that specify independent pre deployment tests, incident disclosure windows, and sandbox escape thresholds without referencing rival coordination. Treat lab safety pledges as context, not a substitute for your own acceptance tests.
What to watch
A published standards framework signed by OpenAI, Anthropic, and Google DeepMind, Microsoft's participation given Frontier Model Forum history, and any Congressional response that converts voluntary bodies into mandatory gates before fall sessions end.
Sources
- Primary. Bloomberg, OpenAI Says It's Working With Anthropic, Google on AI Safety (15 September 2026). Lehane confirmation of multi week coordination and antitrust waiver stance.
- Secondary. CNBC, OpenAI, Google, Anthropic discussing collaboration on AI safety issues (15 September 2026). Hassabis July Standards Body proposal context and Amodei pacing essay linkage.