{"type":"video","version":"1.0","html":"<iframe src=\"https://www.loom.com/embed/3660f6a39efe4f2883b24a4da0e190b4\" frameborder=\"0\" width=\"1114\" height=\"835\" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>","height":835,"width":1114,"provider_name":"Loom","provider_url":"https://www.loom.com","thumbnail_height":835,"thumbnail_width":1114,"thumbnail_url":"https://cdn.loom.com/sessions/thumbnails/3660f6a39efe4f2883b24a4da0e190b4-3949a8339e3b4e55.gif","duration":300.054,"title":"Zero Trust AI Evaluation with Secure Enclave","description":"This Loom explains a zero trust, secure enclave evaluation framework for fairly benchmarking LLM safety across organizations and languages. The speaker describes a process where benchmark owners provide prompts or datasets, model owners provide weights, and independent researchers provide evaluation code, with results returned only after approved code runs inside the enclave. They report evaluating 47 harm prompt cases across four models, noting GPT refused harmful Vietnamese prompts more in English than in Vietnamese, and the most Vietnamese fluent model was least Vietnamese safe. The review concludes that privacy enhancing secure technology enables AI evaluation without requiring trust between parties."}