Anthropic and OpenAI Move to Host Independent Safety Evaluators Inside Their Labs

Anthropic and OpenAI are pushing to embed independent safety evaluators inside their own AI labs, a structural change to how frontier models get scrutinized before release. The arrangement would give outside researchers standing access to models and internal processes that have long been closed to anyone without an employment contract. It marks the first time major labs have proposed handing over that kind of sustained access voluntarily rather than under legal pressure.

The proposal lands as governments in the US, UK, and EU build out formal evaluation regimes. The UK AI Safety Institute has negotiated access agreements with several labs, and the EU AI Act will require systemic-risk model evaluations once its obligations phase in. Against that backdrop, an in-house evaluator program could either complement public oversight or substitute for it, depending on how much autonomy the evaluators actually get.

Researchers who study AI governance say the access on offer is genuinely new. Until now, external safety work has relied on API-level testing, published model cards, or red-team exercises run on the lab's terms. Embedding evaluators inside the building would let them see training data decisions, fine-tuning choices, and deployment plans before a model ships. That is a meaningful upgrade over the status quo, and several researchers quoted in coverage of the plan welcomed it on those grounds.

The objections center on independence. An evaluator paid, hosted, or contractually bound by the lab it is assessing faces obvious conflicts, and past voluntary commitments have drawn criticism for being self-policed. Without guaranteed publication rights, protection from retaliation, and a defined scope of what they can inspect, embedded evaluators risk becoming a credibility layer rather than a check. Critics also note that labs can select which evaluators they invite and which findings see daylight.

What this changes for the field depends on details neither company has fully specified: who appoints the evaluators, who owns their findings, and what happens when a report is unfavorable. If the model holds, it could become a template that regulators point to as evidence of industry cooperation. If it stalls, it may accelerate calls for mandatory third-party audits written into law rather than offered by the labs themselves.

The near-term test is whether either company publishes the terms of the arrangement, including funding, reporting lines, and publication rights. Watch for the first named evaluators and whether their work product is released in full. The open question is simple: an evaluator who cannot publish a negative finding is not an evaluator, and neither lab has yet said what happens when one does.