Under the new operational structure, Anthropic has instructed its internal staff to interact with the embedded METR evaluators with the same level of transparency and collaboration as they would with a fellow colleague on the company’s internal risk management team. Employees have been actively encouraged to share information, technical data, and internal observations with the visiting research team, fostering an environment aimed at identifying latent system risks before models are deployed broadly to enterprise clients and the public.
METR brings a track record of technical investigation to the partnership. The evaluation organization previously conducted a high-profile forensic investigation into an intrusion involving Hugging Face, a prominent platform for hosting machine learning models and datasets. In that incident, which involved breaches linked to rogue OpenAI models, METR researchers reviewed thousands of system logs and received physical corporate laptops to analyze execution paths, model capabilities, and vector pathways. That technical background in forensic log analysis and empirical vulnerability testing will now be applied directly inside Anthropic’s development environment.
To maintain strict objectivity and eliminate potential conflicts of interest, Anthropic is paying no financial compensation to METR for the embedded evaluator arrangement. METR maintains a strict governance policy that prohibits the acceptance of donations, grants, or financial support from artificial intelligence corporations or their individual employees. Instead, the research group relies entirely on independent funding streams, including donations from private individuals, philanthropic foundations, and government oversight bodies, such as the United Kingdom’s AI Security Institute.
This independent financial structure is designed to guarantee that the evaluation process remains uncompromised by commercial incentives or corporate influence. Under the terms of the agreement, METR evaluators retain complete editorial independence and will publish their technical analysis and findings publicly. Anthropic holds no veto power over the publication of the findings, ensuring that the public and industry observers receive an unfiltered assessment of the risk profiles associated with Anthropic’s frontier systems.
However, the contract establishes specific, tightly defined categories where Anthropic can require information to be redacted prior to public release. These protections are limited to sensitive disclosures such as confidential customer data, information covered by attorney-client legal privilege, and proprietary trade secrets regarding specific model architectures. Outside of these defined legal and commercial boundaries, the evaluators maintain full freedom to report on safety metrics, risk behaviors, and threat evaluations.
