1/ In an announcement dated 2026-10-15, Anthropic said it has begun granting employee-level access to METR, a nonprofit evaluation organization, to better assess the safety of its systems.
2/ The company explained that this measure is intended to allow external evaluators to inspect systems under conditions closer to a real working environment.
3/ Anthropic also added that it is in early discussions with other organizations, including Hugging Face.
4/ This announcement shows a trend toward institutionalizing independent external evaluation rather than limiting checks to internal review within the model developer.
5/ In particular, it reveals a direction of linking safety verification not to one-off reviews but to continuous operational procedures. [1]
Implications
6/ This move suggests that AI safety verification may shift from internal testing to external independent evaluation.
7/ For researchers, it is important that access and evaluation frameworks close to real usage environments can more precisely reveal model behavior, misuse potential, and the risks of agentic functions.
8/ For the materials and science research community, this can be read as a message that when connecting models to laboratory workflows for tool use, automation, and long-running tasks, one should design not only for performance but also for auditability, access control, and independent evaluability.
9/ In the future, external verification procedures may become standardized on platforms handling high-risk functions. [1]