A researcher discovered 6TB of exposed data from a third-party AI model router, containing sensitive interactions with models like Claude. The leak demonstrates how intermediate 'routers' used to manage AI traffic are becoming a major security vulnerability for users not running models locally.