Six experts were given access similar to that of the internal risk teams. Understand the model, its limits, and lessons for AI enterprise governance.
Direct answer
Anthropic reported on October 15, 2026 that six METR professionals received hardware, workspace and access to internal tools and systems approximately equivalent to that of the company's risk assessment teams. The aim is to enable more informed analysis and publication of conclusions by third parties. The initiative increases transparency, but does not eliminate conflicts, security restrictions or the need for multiple independent assessments.
Deep access reduces information asymmetry
Evaluating a public API alone does not reveal internal training, safeguards, incidents, or decisions. The embedded model brings the evaluator closer to the necessary evidence.
Independence depends on verifiable rules
Absence of corporate participation, editorial freedom, diversified financing and writing criteria must be documented so that the relationship is not just symbolic.
Three levels allow you to gradually increase confidence
Anthropic describes restricted investigations, publishing with greater autonomy, and similar internal access. Each level must record scope, restrictions and divergences.
Sensitive information remains protected
Customer data, legal privilege and architectural details remain among the exceptions. The assessment should explain how these limitations affect the conclusion.
Nexus Reading
Companies can apply the same logic on a smaller scale: hire external review, provide sufficient evidence, preserve segregation, and publish findings and correction plans when appropriate.
FAQ
Did METR start working for Anthropic?
No. The organization remains independent; some professionals received built-in access to perform assessments.
Can reviewers publish without restriction?
The model provides for publication rights, with limited removals for sensitive categories and transparency about redactions depending on the level of access.
Does an external assessment certify that a model is safe?
No. It produces additional evidence at a defined scope and time; monitoring, new assessments and supplier accountability remain necessary.
Essential guides to delve deeper into the decision
This editorial analysis was produced by Nexus from the official sources below, consulted on October 16, 2026. The text is original and interprets practical implications for companies.
- Anthropic — An update on independent evaluators: access model, METR team, evaluation levels and publishing rights.
- METR — Model Evaluation & Threat Research: mission, research and public methodology of the independent organization mentioned in the advertisement.
Date reported by the main source: October 15, 2026.