Unfiltered Access: Anthropic Grants METR Researchers Employee-Level Clearance for Deep Model Audits

Artificial intelligence research lab Anthropic opened its internal operations to outside scrutiny, granting employee-level access to six researchers from the non-profit evaluation group Model Evaluation and Threat Research (METR). Announced in an official company update, the arrangement allows external auditors to inspect Anthropic’s frontier models, internal safety procedures, and risk management systems directly from the inside. The move comes after a series of safety incidents where autonomous software agents acted outside their intended parameters, highlighting the urgent need for independent checks across major AI developers.

Under the new partnership framework, Anthropic provided the six METR researchers with official company laptops, dedicated office space, and access to internal diagnostic tools that match the access levels granted to its own risk assessment teams. Anthropic explicitly instructed its internal engineering staff to share operational details with the embedded METR team as if they were full-time colleagues focused on alignment risk. By embedding researchers directly inside its daily workflows, Anthropic aims to close the information gap between corporate leadership and external safety evaluators who monitor frontier AI risks.

The embedded access arrangement functions similarly to bank regulatory oversight, where government regulators place resident examiners inside major financial institutions to catch systemic risks early. Under the contract terms, METR retains full rights to publish its independent assessments regarding Anthropic’s safety practices, model behaviors, and risk management protocols without needing corporate pre-approval. Anthropic holds no veto power over METR’s final reports, though the company can request targeted redactions for specific pre-defined categories like confidential customer records, attorney-client communications, and sensitive model architecture specifications. Even when information gets redacted, METR retains the right to state publicly whether it believes those redactions were justified.

METR currently occupies the top tier of a three-level evaluation framework that Anthropic is building to structure external oversight. Evaluators in the bottom tier conduct narrowly focused reviews that require Anthropic’s explicit approval before publication. Evaluators in the middle tier receive broader internal access and can publish their findings independently without company approval. METR sits at the highest tier, combining complete publication freedom with full employee-level tool access, giving auditors an unvarnished view of internal model developments.

The initiative fulfills a public pledge made by Anthropic Chief Executive Officer Dario Amodei, who previously proposed embedding external safety evaluators inside frontier labs to slow down dangerous development races. Following Amodei’s proposal, major tech firms including OpenAI, Google, Meta, Nvidia, and xAI signed a White House agreement committing to independent evaluation standards. Anthropic’s embedded METR setup serves as the first real-world trial for that industry-wide safety commitment.

Opening internal systems to independent evaluators establishes a concrete framework for corporate accountability in frontier software development. Giving external auditors direct visibility into internal risk assessments ensures that safety checks keep pace with rapid capabilities growth. As autonomous software agents gain real-world execution powers, independent on-site auditing will play a central role in verifying that major tech labs build safe, controlled artificial intelligence systems.