1/ On 2026년 10월 15일, Anthropic announced that it had begun granting employee-level access to the independent evaluation organization METR.
2/ The company explained that this is intended to enable stricter verification of its frontier models and safety practices.
3/ The key point of this measure is that it expands the scope of external evaluation so that it is carried out not as a simple post hoc check, but in a manner closer to internal access.
4/ Anthropic also said that, in addition to METR, it is conducting evaluation work with SecureBio and Accenture, and is continuing additional discussions with Hugging Face and others.
5/ This shows a shift away from verification by the model developer alone toward a direction in which external actors in the security, life sciences, and platform ecosystem take a deeper role in safety evaluation. [1]
Implications
6/ Granting independent evaluators internal-level access may increase the trustworthiness and reproducibility of safety verification.
7/ In particular, when agentic LLMs are used in materials and scientific research for experimental design, automation, literature search, and data processing, strengthened external verification systems may further specify the criteria for judging the reliability of research tools.
8/ In addition, in research infrastructure where security and access control are important, the evaluation and audit systems of model providers may directly affect adoption criteria. [1]