Anthropic METR Access Grants Employee-Level Permissions for Independent AI Safety Evaluation

On October 15, 2026, Anthropic announced it has granted six METR employees employee-like access to its internal systems, hardware, and office space for independent AI safety evaluations. The arrangement places the evaluators on equal footing with Anthropic’s own risk teams outside narrow protected categories such as customer data and model architecture details.

Anthropic METR Access Details Under the New Tiered Evaluator Framework

The six embedded assessors receive company hardware and tools comparable to internal staff. Anthropic staff are instructed to share information with the METR team as they would with colleagues focused on alignment and risk work. METR retains full rights to publish its findings on risk levels, incidents, and practices, while Anthropic may request redactions only in predefined areas.

METR operates under strict siloing agreements that separate this embedded group from its other researchers. The nonprofit receives no compensation from Anthropic. This setup forms the third and highest tier of Anthropic’s evaluator framework, where only METR currently holds employee-level privileges. Lower tiers allow narrower reviews or publication without prior approval but with less access.

Anthropic also maintains partnerships with SecureBio and Accenture for additional reviews. The company previously signed the White House Accord on Super Intelligence and committed to expanded third-party assessment. METR has conducted earlier risk report reviews with Anthropic and other labs without financial ties.

Background on Independent Oversight for Frontier Model Risk

Frontier model developers have faced repeated calls for stronger external scrutiny as model capabilities advance. Anthropic positions the embedded access model after the structure of bank examiners who receive detailed internal data while maintaining independence. The company states the goal is to close information gaps that have limited prior external evaluations.

Under the framework, Tier 1 covers narrowly scoped reviews requiring Anthropic approval for publication, as seen with SecureBio’s work on the chemical and biological weapons section of the August 2026 Risk Report. Tier 2 grants broader access and independent publication rights subject to redactions. Tier 3, now active with METR, adds day-to-day integration with internal systems.

Early discussions involve additional organizations including Hugging Face. Anthropic plans to release periodic updates on the partnerships but has not announced a timeline for METR’s first public reports from the embedded arrangement. The structure aims to balance transparency with protections for sensitive information.

Practical Implications for AI Safety Evaluation and Oversight

Organizations tracking frontier model risks gain a clearer view of internal processes through METR’s independent publications. Readers can compare METR’s assessments against Anthropic’s own reports to identify discrepancies in risk characterization. The arrangement provides a concrete example of how labs can structure deeper external review without ceding control over proprietary details.

Developers and investors evaluating AI governance practices now have a documented benchmark for third-party access levels. The METR model demonstrates enforceable separation between embedded evaluators and the rest of the nonprofit, addressing concerns about information leakage. Companies considering similar programs can reference the three-tier structure when designing their own oversight mechanisms.

Future reports from the embedded METR team will offer ongoing data points on Anthropic’s risk management effectiveness. Stakeholders monitoring AI safety evaluation standards can track how redactions and scope limitations evolve over successive publications.