{"type":"video","version":"1.0","html":"<iframe src=\"https://www.loom.com/embed/281588dc362f4c2bb92f4b7185ee9300\" 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/281588dc362f4c2bb92f4b7185ee9300-a425e6407b7c08c9.gif","duration":106.92,"title":"Using Airlock to Prevent Data Exfiltration","description":"This Loom explains how Airlock prevents AI data exfiltration when an agent is prompted with sensitive files. Shifman demonstrates how asking an AI, via a codex and a request like finding whose birthdays are in the next 13 days, could leak employee census data including date of birth and SSNs. He then shows that using Airlock protects the file by replacing direct reading with requests to a context broker that enforces policies, such as not revealing exact data like specific ages. The example includes narrowing results based on answers about the keynote, leading to names like Maya and Jordan, but blocking exact age details."}