Microsoft Uncovers 'Whisper Leak' Attack That Identifies AI Chat Topics in Encrypted Traffic - The Hacker News

The "Whisper Leak" is a novel side-channel attack targeting remote language models, allowing passive adversaries to infer sensitive conversation topics from encrypted network traffic. This is achieved by analyzing packet sizes and inter-arrival times of streaming LLM responses, which enables trained classifiers to reliably identify specific prompt topics, posing a significant privacy risk.