Anthropic METR Access Grants Employee-Level Tools for Frontier Model Risk Evaluation

On October 15, 2026, Anthropic provided a six-person team from METR with employee-like access to its internal systems, hardware, office space, and tools. The move supports independent evaluation of frontier AI model risks and follows the company's three-tier evaluator framework.

Details of the Tier 3 Access for METR

The access level matches what Anthropic grants its own risk assessment teams. METR researchers can interact directly with staff on alignment risks, subject to narrow limits on customer data, attorney-client materials, and model architecture details. This setup represents the highest tier in Anthropic’s new framework, where Tier 1 covers scoped reviews with approval requirements, Tier 2 allows publication after redactions, and Tier 3 adds unrestricted access plus strong publication rights.

METR will issue its own reports on risk levels, incidents, and safety practices. Anthropic may request redactions only in defined categories such as confidential intellectual property or customer data, but it cannot block publication. The nonprofit may also comment publicly on the scope of information received. The six-person team works under enforceable agreements and remains siloed from other METR personnel.

Anthropic does not fund METR. The organization draws support from philanthropic sources and government bodies including the UK AI Security Institute. This independence distinguishes the arrangement from paid consulting relationships.

Background and Industry Context

METR has conducted prior reviews of Anthropic risk reports and capability evaluations, alongside similar work with OpenAI, Google, and Meta. The expanded access follows Anthropic CEO Dario Amodei’s September 2026 comments on pacing frontier development and commitments under the White House Accord on Super Intelligence. Anthropic is already running Tier 1 work with SecureBio on the chem/bio section of its August 2026 Risk Report and maintains an embedded evaluation partnership with Accenture and Faculty.

Readers in AI safety, investment, and engineering roles gain a concrete example of how independent evaluators can operate inside a frontier lab without compromising core protections. The arrangement supplies verifiable data on model behaviors that internal reports alone cannot fully address, allowing external analysts to cross-check claims about capability thresholds and mitigation measures.

Early discussions with additional organizations such as Hugging Face indicate Anthropic intends to extend Tier 3 access beyond METR. METR’s forthcoming assessments will provide the first public test of the model’s effectiveness in practice.

Next Steps and Practical Implications

METR plans to publish risk assessments based on the new access. Anthropic has stated it will share further updates as additional evaluators reach Tier 3 status. Organizations evaluating AI suppliers or building internal oversight programs can use the published METR findings to benchmark their own review processes against an independent standard.

The structure also offers a template for other labs weighing similar embedded evaluator programs. By separating funding from access rights and limiting redactions to narrow categories, the approach reduces conflicts of interest while preserving legitimate commercial protections. Professionals tracking AI governance developments should monitor the first METR reports for concrete metrics on how the access translates into measurable evaluation depth.