{"type":"video","version":"1.0","html":"<iframe src=\"https://www.loom.com/embed/1c1476293fce47b7bf06af1fef57bed3\" frameborder=\"0\" width=\"1318\" height=\"988\" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>","height":988,"width":1318,"provider_name":"Loom","provider_url":"https://www.loom.com","thumbnail_height":988,"thumbnail_width":1318,"thumbnail_url":"https://cdn.loom.com/sessions/thumbnails/1c1476293fce47b7bf06af1fef57bed3-ce5ce60f94bcf092.gif","duration":291.104,"title":"Clinical Co Pilot Security Red Team Check-In","description":"This Loom discusses the week three security check-in for the Clinical Co-Pilot Adversarial Agent and its threat-focused testing approach. It explains how authenticated users may still be able to trigger failures, such as via malicious text in uploaded intake documents, and describes a four-agent adversary that simulates user inputs to improve reliability over time. In two runs, the second run produced an additional failure due to a deterministic judge false positive when the agent correctly refused to access allergies for a patient named Marcus Hill but still flagged the response because the name appeared. The Loom then outlines the iterative lifecycle for evaluating, documenting, fixing, and running regression until vulnerabilities are resolved, with full automation targeted for Friday’s deployment."}