Anthropic reportedly brought six Model Evaluation and Threat Research (METR) employees into its offices to assess risks from Claude models. On October 15, 2026, the company said the evaluators would receive company hardware, office space, and access to internal tools and systems roughly comparable to Anthropic’s risk-assessment teams, according to Ars Technica’s republication of Cristina Criddle’s Financial Times report.
The arrangement also gives METR the right to publish its analysis without Anthropic’s approval, with redactions allowed for specified sensitive information. By October 18, no completed METR findings from the embedded evaluation had been reported.
What METR could access
The six evaluators received company hardware and office space, along with access to internal tools and systems described as roughly on par with those available to Anthropic’s risk-assessment teams. Anthropic also encouraged employees to share information with the METR evaluators as they would with a colleague working in risk.
What METR could publish
METR could publish its analysis without seeking Anthropic’s approval. Anthropic could require redactions for customer data, attorney-client privileged information, and details about model architecture. The publication right therefore came with defined limits on what could appear in a public report.
How the funding relationship was described
The report said Anthropic was not paying METR for the arrangement. METR also does not accept donations from AI companies or their employees. Its support comes from individuals, philanthropic organizations, and government bodies, including the UK AI Security Institute.
The arrangement has no reported evaluation result
The October 18 coverage described the access and publication terms, but no completed METR findings or pass/fail result for Claude had been reported by that date.
