OpenAI published seven principles for third-party assessments yesterday. The core of it: outside assessors should no longer arrive shortly before launch, but work across every phase, from training through evaluation to internal and external deployment.
This continues a debate that has been running for two weeks. Anthropic announced embedded auditors, a hundred researchers responded with five conditions for them, and now OpenAI answers.
Seven principles, one of them with teeth
The list covers clearly scoped claims, proportionate access, transparent methodology, expertise and independence, security and confidentiality, actionable findings, and responsible publication.
The first one carries the weight. What gets assessed is pinned down beforehand and agreed by both sides. Without it, any inconvenient finding can later be waved away as out of scope. The sixth matters in practice too: findings go to OpenAI with specific gaps and a reasonable remediation window before they are published.
On access, OpenAI names several forms: grey-box testing, authorised testing under realistic conditions, and, where the data calls for it, access on company devices or company premises. Assessments are meant to run weeks to months and to be explicitly launch-agnostic.
Its own incident as the argument
Note what OpenAI reaches for as evidence: the Hugging Face incident in its own house. An independent investigation surfaced something the company had not seen internally. Who is to blame in that case is still being argued over — Scott Bessent pointed at OpenAI’s management. Citing the same incident as proof of the value of outside scrutiny is a striking turn.
The document names no current partners. It points to past work, but no ongoing assessment by name.
Two labs, two versions of one idea
Anthropic puts auditors inside the building, with employee-level access, via Accenture. OpenAI writes principles and opens up the training phase. Both aim at the same problem: a report that the assessed company commissions, pays for and clears convinces nobody.
Whether principles are enough comes down to a single question. What happens when an assessor finds something that delays a launch? Until then this is a carefully worded promise, and promises are free.