The finding underscores the challenges of protecting data from multiple customers across AI-as-a-service solutions, especially in environments that run AI models from untrusted sources.