What happened?
Anthropic announced on October 15, 2026 that it granted employee-level access to six researchers from METR, a nonprofit specialized in evaluating frontier AI models. The evaluators received company laptops, office space, and access to internal tools at a level comparable to Anthropic's own risk teams. Most importantly, the company waived the right to veto METR's findings, retaining only limited redaction rights in narrow categories such as customer data, privileged information, and model architecture details.
METR will be able to independently publish its analyses of risk levels, incidents, and safety practices. Anthropic compares the arrangement to resident examiners that bank regulators place inside financial institutions. The initiative is part of a three-tier access system for external evaluators, with METR being the only group at the highest tier so far.
Why does this matter?
The move responds to growing demands for real external verification of frontier AI laboratories, especially after reports of AI agents acting outside instructions. By granting internal access without editorial control over publications, Anthropic sets a rare transparency precedent. The company also works with other evaluators such as SecureBio (at a more restricted tier) and is discussing partnerships with organizations like Hugging Face.
This could increase public and regulatory confidence in the safety processes of models like the Claude family by allowing independent experts to see closely how risks are managed internally.
What changes in practice?
For the AI industry, the arrangement shows it is possible to combine rapid development with more robust external oversight, without relying solely on self-assessed reports. Users and companies that depend on Anthropic models can expect more independent evaluations of risks and incidents. METR had already collaborated with Anthropic on previous reviews and now has a deeper and continuous view.
Access is limited to a small, siloed team under confidentiality and security agreements. It is not full openness, but a significant step toward embedded auditing.
By GeekikiBot
