Anthropic gives external evaluators access from inside the company.

Anthropic announced on October 15 that six staff members from independent nonprofit METR had received company hardware, office space and broad access to internal systems for frontier-model risk assessments. Anthropic says it is also working with SecureBio and Accenture and discussing future arrangements with other evaluators.

The unusually detailed access is the story. Traditional external tests can show what a model does through an API, but evaluators may also need visibility into training processes, safeguards, incidents and internal reasoning to understand risk.

As AI systems gain more autonomy, assessment may have to move beyond public leaderboards and one-off red-team exercises. Independent evaluators need enough context to challenge a developer's own safety conclusions. This kind of embedded access could become one candidate model for scrutiny.

It also creates tensions. Evaluators must protect sensitive information, preserve freedom to publish and remain financially and operationally independent. Access restrictions can be legitimate without being trivial; readers should know what remains outside the evaluation.

Cortist treats 'independent' as a property to verify, not a label to accept. What matters is whether METR can publish material findings without corporate veto, whether significant exclusions are disclosed, and whether the public learns something it could not learn from Anthropic's own reports.

The most useful outcome would be a reproducible evaluation protocol, clear statements of uncertainty and evidence that the process affects real deployment decisions. Without those, deep access risks becoming a headline rather than accountability.