Anthropic announced on October 15, 2026, that it had begun onboarding six employees of the nonprofit Model Evaluation and Threat Research (METR) with access to internal tools and systems roughly on par with the company’s risk-assessment teams. The next day, Anthropic clarified that it decides whether information fits its redaction categories, while METR may make certain statements about the access and information it received. Anthropic’s announcement describes the arrangement as the third tier in a three-level evaluator framework.
What access does METR receive?
Anthropic says the six-person team has company hardware and office space, as well as employee-like access to internal systems. The team operates under enforceable agreements. Other METR employees—including researchers working with Anthropic on separate projects—are siloed from its work.
That makes this a defined team arrangement, not access for METR as a whole. Anthropic describes the evaluators’ purpose as assessing risks from frontier models and the company’s safety practices.
How Anthropic’s three evaluator tiers differ
Anthropic’s framework sets out different access and publication terms for each tier. METR’s arrangement is in tier three.
| Tier | Access scope | Publication terms |
| One | Narrowly scoped risk investigations; Anthropic decides what access to provide. | Anthropic approval is required before evaluators publish findings. |
| Two | More information than tier one. | Evaluators may publish without prior approval, subject to specified redactions, and may publish their views on whether redactions were justified. |
| Three (METR) | Employee-like access to internal tools and systems, roughly on par with Anthropic’s risk-assessment teams. | Publication rights similar to tier two. |
Who decides what METR can publish?
Anthropic says it cannot veto METR’s publication, but it can require redactions for predefined categories. The company also says it has the final say on whether information falls into those categories. METR, in turn, may make certain statements about the level of access and information it received.
Anthropic names customer information, attorney-client privileged information, and model-architecture information as examples of categories that may be withheld. It says architecture information may be shared when highly relevant, generally with explicit internal approval. So publication is permitted under the arrangement, but Anthropic retains a role in deciding what information can appear.
The question of METR’s independence has a longer backstory. NeoTeo previously examined questions surrounding METR and Anthropic’s oversight plan. In its announcement, Anthropic says evaluators should have expertise in assessing frontier models, be independent of the company they assess—including having no equity or financial or operational dependence—and have enough access and freedom to publish observations.
What other evaluator work is Anthropic pursuing?
Anthropic says SecureBio is reviewing the chemical-and-biological-weapons chapter of its August 2026 Risk Report, after completing and publishing a pilot review of an earlier report. Separately, Anthropic says it is working with Accenture and Faculty toward a partnership covering model evaluation and red-teaming, alignment assessments, safeguard testing, and capability-building.
Anthropic also says it is in early discussions with other organizations, including Hugging Face.
