Microsoft has discovered a side-channel attack on language models which allows adversaries to conclude model conversation topics, despite being encrypted. The post ​​Whisper Leak: A novel side-channel attack on remote language models appeared first on Microsoft Security Blog .