Anthropic Embeds Independent Safety Team With Publishing Rights

What Happened

Anthropic has formally embedded a team from METR, an independent AI safety research organisation, inside its laboratory and granted that team the right to publish its findings without requiring Anthropic's approval. The arrangement gives external researchers direct access to Anthropic's models and internal processes, with a structural guarantee that their conclusions cannot be suppressed by the company.

Who Is Involved

Anthropic, founded in 2021 and headquartered in San Francisco, develops the Claude family of large language models and has positioned AI safety research as a central part of its mission. METR, which stands for Model Evaluation and Threat Research, is an independent nonprofit organisation focused on evaluating the capabilities and risks of advanced AI systems. METR has previously conducted evaluations for other AI developers, including assessments of autonomous task completion and self-replication risk in frontier models.

Terms of the Arrangement

Under the agreement, METR researchers are embedded within Anthropic's facilities and have access to models and data relevant to safety evaluation. The publishing rights granted to METR are described as veto-free, meaning Anthropic retains no formal power to block, delay, or require changes to METR's public reports. The structure is notable because most external audits or red-teaming arrangements in the AI industry are governed by non-disclosure agreements or publication review clauses that give the commissioning company some degree of control over what is released.

Background

The announcement follows a period of heightened scrutiny of how AI laboratories manage safety testing and communicate risk findings to the public. Anthropic's own internal research, published earlier this year, indicated that some smaller AI models are approaching the capability threshold for self-replication, a finding that drew significant attention from researchers and policymakers. Anthropic's revenue has also grown substantially, with figures reported at a 45 percent increase, which has brought additional attention to how the company balances commercial growth with its stated safety commitments.

The broader AI industry has faced repeated questions about the independence and transparency of safety evaluations. Several major laboratories conduct evaluations internally or through contracted third parties, arrangements that critics have argued may not produce fully independent results. Government bodies in the United States, the United Kingdom, and the European Union have each begun developing frameworks that would require or incentivise independent third-party evaluation of frontier AI systems.

What It Means in Practice

The practical effect of the arrangement is that METR can document and publicly disclose capability findings, risk assessments, or safety-relevant observations about Anthropic's models without the company having a contractual mechanism to prevent publication. This differs from advisory or collaborative research agreements in which the host company retains editorial rights.

METR has not disclosed a specific timeline or publication schedule for reports produced under the arrangement. Anthropic has not indicated whether similar agreements are planned with other external organisations.

Industry Context

Several governments and intergovernmental bodies have cited independent evaluation capacity as a prerequisite for effective AI governance. The United Kingdom's AI Safety Institute and the United States AI Safety Institute, established in 2023 and 2024 respectively, have each worked to develop evaluation protocols and to negotiate access agreements with major AI developers. Those agreements have varied in scope and in the degree of publication independence they provide to government evaluators.

The METR arrangement at Anthropic represents one of the more structurally defined examples of veto-free third-party access at a leading AI laboratory, though the full scope of model access and evaluation scope has not been publicly detailed.

METR is expected to publish findings from its work inside Anthropic on a schedule the organisation determines independently.