​​Whisper Leak: A novel side-channel attack on remote language models

Microsoft researchers disclose a novel side-channel vulnerability (Whisper Leak) against streaming-mode language models that allows an observer of encrypted network traffic to infer whether a user’s prompt concerns a specific sensitive topic by analyzing packet sizes and timings. The report describes the experimental methodology, machine-learning classifiers that achieved high precision in controlled tests, real-world impact simulations, and coordinated mitigations (obfuscation/response padding) deployed with cloud providers to reduce practical risk.