Anthropic gives a six-person METR team employee-like access and no veto over publishing its risk assessments

Core Background & Industry Pain Points

Frontier labs have largely graded their own risk: training details and safeguard performance are commercially sensitive or misuse-prone, so outsiders only see curated reports. Last month Anthropic committed to stronger third-party assessment and signed the White House Accord on Super Intelligence, pledging to work with independent auditors or evaluators. The METR arrangement is the first concrete step.

Architecture Highlights & Mechanism

Anthropic compares it to resident bank examiners. Six METR employees get company hardware, office space and system access roughly on par with internal risk teams, and staff are encouraged to share information as with a colleague, except for a few categories such as customer data, privileged legal material and model architecture. METR may publish its analysis of Anthropic's risk levels, incidents and practices; Anthropic cannot veto it and may only require redactions in predefined categories (Anthropic decides what falls in them, while METR may describe the access it received). The embedded team is siloed from other METR staff. Evaluator access has three tiers, with METR at tier 3.

Benchmarks & Measured Results

This is a governance and evaluation update with no model scores; no new benchmarks were published. For progress: SecureBio already published a pilot review of the previous risk report and is now reviewing the chem/bio chapter of the August Risk Report, while the Accenture and Faculty partnership covers red-teaming, alignment assessments and safeguard testing.

Developer Adoption Guide

Teams building on Claude should watch for METR's independent assessments, which will be more useful than vendor self-reports in procurement and vendor-risk reviews. Read Anthropic's post and METR's risk assessment page to understand the tier boundaries, and mark in compliance docs which conclusions come from third parties versus the vendor.