Anthropic gives METR deep internal access: What independent AI evaluators can really see
Six METR staff receive employee-like access at Anthropic. Publication rights, confidentiality limits, and what this means for frontier-agent safety.
On October 15, 2026, Anthropic said it had onboarded six staff from the independent research organization METR with employee-like access to internal systems. METR is expected to evaluate risks and safety practices and publish its own findings. Important exceptions for confidential information and redaction authority remain.
External evaluators typically assessed frontier systems with narrower information or access. Visibility into internal risk decisions, incidents, and safeguards was limited.
A small METR team now has company hardware, office space, and internal tools roughly comparable to Anthropic's own risk teams, with publication rights and defined confidentiality boundaries.
What employee-like access actually means
Anthropic says six METR employees have been onboarded with corporate hardware, workspace, and access to internal tools and information broadly comparable to internal risk-assessment staff. Anthropic employees are encouraged to share relevant information with them as they would with colleagues evaluating model risks. That is materially different from providing an API endpoint for external benchmarking or permitting a short, supervised demonstration.
There are clear exceptions. Anthropic identifies customer data, legally privileged communications, and some model architecture details as protected categories. The embedded arrangement applies to a small METR team under enforceable agreements, with other METR researchers siloed from that work. Such protections may serve legitimate security and privacy purposes, but they also affect what outside assessors can inspect directly.
Independent publication, but limited redaction power
METR is expected to publish analyses of risk levels, incidents, and safety practices. Anthropic states that it cannot simply veto publication, but may require removal of information in predefined protected categories. Crucially, Anthropic ultimately determines what falls into those categories; METR may still disclose how much access it received and whether it believes redactions were justified.
These details matter more than an abstract claim of independence. Readers should ask what evidence evaluators could see, whether they could test competing explanations, and whether difficult findings make it into public reports. An access agreement announced today is not a completed positive audit, and it does not certify that Claude or agentic deployments are universally safe.
A three-tier framework for further evaluations
Anthropic describes three tiers of evaluator access. At the first tier, investigations are narrowly scoped and publication requires vendor approval. The second offers more publication freedom but less information than an embedded team. The third is the METR-style arrangement: employee-like access alongside the stronger publication rights. Anthropic also names SecureBio and Accenture as other evaluators and says additional partnerships are under discussion.
For organizations deploying Claude-powered agents, this is governance news rather than a new agent feature. Useful next evidence will be independently published assessments, disclosure of their limits, and concrete changes made in response. Partnership announcements should not be treated as proof that autonomous behavior is now comprehensively controlled or that every material risk has already been examined.
AgentenCode assessment
AgentenCode assessment: The meaningful advance is deeper access for independent risk reviewers plus explicit rights to publish. Credibility will depend on what reviewers can verify, how redactions are handled, and whether findings change real safeguards. Beginning an evaluation is not the same as receiving a safety certification.
Sources and verification
Checked October 19, 2026. Anthropic's October 15 primary announcement describes the METR arrangement, excluded information, and three access tiers. It is an announcement of evaluation access, not the publication of a completed independent assessment.