OpenAI opens training-phase safety evals to outsiders

"Independently safety tested" is about to mean something different, and you should find out what. On September 22, OpenAI published priorities and principles for third-party assessments, saying outside organizations will be able to run technical safety assessments during training, evaluation and deployment — not only in the window just before a model ships. For anyone doing AI vendor diligence, that phrase is now a question with a real answer behind it.

What actually happened

Per Bloomberg, the shift moves external evaluators upstream. Until now the standard industry practice — OpenAI's included — has been to bring outside groups in shortly before release, evaluate the finished artifact, and publish alongside the launch.

OpenAI named four priorities it says make such assessments work: independence mechanisms, scientific rigor, security practices, and clear responsibilities.

What it did not name is a partner. TNW reports that OpenAI is in talks with groups including METR and Redwood Research, but no organization is confirmed, no access terms are published, and no timeline is set. TNW also reads the post as narrower than Sam Altman's comments ten days earlier, which described evaluators getting desks, badges and laptops; the published version says evaluators "may be brought into the offices for the most sensitive work."

So: a real direction, stated at a level of specificity that commits nothing yet. Read it as the opening position, not the deal.

Why AI vendor diligence matters for your business

You are not going to audit a frontier lab. But you are going to sign something that says a model is safe for your use, and the meaning of that sentence is being renegotiated in public right now.

The reason it matters for a small operator is that model safety claims flow downhill into your contracts. Your client asks whether their data is safe with your automation. You ask your vendor. Your vendor points at an eval. If that eval was run by the vendor on the vendor's model against the vendor's threat list, you are passing along marketing as assurance — and you are the one in the chain with a signature on it.

Concrete moves:

Ask who ran the eval, and whether they could publish a bad result. Independence is not a vibe. It is a contract term about who pays, who controls the data, and who holds publication rights. If the assessor cannot publish an unflattering finding without the lab's approval, the finding is a press release.

Ask what was assessed, not just that it was. A safety card covering refusal behavior tells you nothing about tool-use scope or credential handling — which, as this week's Gemini disclosure showed, is where agents actually fail.

Put model-change notice in the contract. Training-phase evaluation is only useful if you know when the thing you evaluated got replaced. Ask for notice before a model version changes under an API alias, and pin versions where your vendor allows it.

Do not outsource your own controls. Upstream evals reduce the odds of a catastrophic model. They do not scope your API tokens, cap your spend, or log what your agent touched. Those are yours regardless of who audited what.

Sources: OpenAI, Bloomberg, TNW.