Microsoft Uncovers 'Whisper Leak' Attack That Identifies AI Chat Topics in Encrypted Traffic
Summary: Microsoft disclosed a side-channel attack it calls Whisper Leak, in which a passive adversary observing encrypted TLS traffic between a user and a streaming LLM can extract packet size and timing sequences and use trained classifiers to infer whether the conversation topic matches a sensitive category. The researchers trained a binary classifier using LightGBM, Bi-LSTM, and BERT and found models from Alibaba, DeepSeek, Mistral, Microsoft, OpenAI, and xAI scoring above 98%, while Google and Amazon models showed greater resistance, likely due to token batching.
Key Takeaway: Microsoft notes that a government agency or ISP monitoring traffic to a popular AI chatbot could reliably identify users asking about topics such as money laundering or political dissent even though the traffic is encrypted, and that attacker success rates can rise with more training samples and multi-turn conversations. OpenAI, Mistral, Microsoft, and xAI have deployed mitigations, including adding a random sequence of text of variable length to each response to mask token lengths.
Read the original at The Hacker News ↗
- Topics
- Privacy
- Related
- Italy Fines Meta €805 Million Over Facebook's Android Gallery Access Without Explicit Consent Cyber Risks · Nov 15, 2026