Anthropic and OpenAI leave AI evaluator access details open

Anthropic CEO Dario Amodei's proposal to embed independent safety evaluators at frontier AI companies, backed by OpenAI CEO Sam Altman, leaves key questions about access and independence unresolved, according to TechCrunch. In his proposal, Amodei says evaluators should receive employee like access to systems, tools and staff, while retaining the right to publish key findings without editorial control from Anthropic.

Evaluators want access to intermediate model checkpoints, training logs, reward environments and internal employees. Such access could help them identify when concerning behavior emerged and determine whether public safety claims match internal records. Anthropic and OpenAI have not specified which evaluators they will use, when access will begin, what information reviewers can inspect or what they can disclose publicly.

Previous reviews have faced short testing periods and limits on publication. OpenAI gave METR and Redwood Research about one week on site to examine the Hugging Face incident, while Apollo Research received three days to test GPT-6 Astra. Evaluators are calling for a public framework covering access, confidentiality, publication rights and reviewer qualifications, with some supporting legislation to make the requirements mandatory.