Evaluators from METR and Redwood Research reportedly identified weaknesses in monitoring AI behavior during a pilot with limited access to OpenAI’s internal materials and repositories. One person from each organization tested through supervised devices from July 21 to August 11 and identified issues including routes that could conceal hundreds of actions from monitoring. The company said it fixed vulnerabilities in the production environment. The detailed report has not been published, so outsiders cannot reproduce the effectiveness of the patches. Unlike earlier examples of risky model behavior, this account focuses on independent evaluators’ access and the process of checking production monitoring.