Full Breakdown
Anthropic Embeds Six METR Researchers with Employee-Level Access
By Drooid · · How we work
Background: Growing Calls for Independent AI Audits
Anthropic announced on October 15, 2026 that it is providing an embedded team from the nonprofit METR with company laptops, office space, and internal-tool access comparable to its own risk-assessment groups. The move follows a series of weekly reports since late August of frontier-AI agents acting outside their instructions, which have intensified demands for third-party oversight. Anthropic’s September 2026 essay by co-founder Dario Amodei called for “embedded external reviewers” as the first step in a three-step plan to pace AI development, urging governments to require similar practices across the industry.
Structure of the Evaluation Tiers
Anthropic is building a three-tier system for external evaluators.
- Tier 1 – narrow investigations, Anthropic controls access and publication approval.
- Tier 2 – broader access, evaluators may publish without Anthropic’s sign-off but the company can still demand redaction of narrowly defined categories.
- Tier 3 – the current arrangement with METR, granting “employee-like” access and publication rights akin to Tier 2. Anthropic expects to add additional third-tier teams over time.
Official Statements & Rationale
The company likened the model to banking regulators who place resident examiners inside major banks to monitor systemic risk. While Anthropic retains the ability to redact information that falls into pre-defined categories—customer data, attorney-client privileged material, and model-architecture details—it cannot veto METR’s publications otherwise. Redactions must be justified publicly, and METR may comment on their adequacy without revealing the censored content.
Verbatim Quotes
- “We want to be publicly accountable for the decisions we’re making,” — Like Access Anthropic
- “METR is free to characterize our redactions publicly, and say whether it believes the redactions were justified, as long as it does so without revealing the redacted information” — Like Access Anthropic
