Whisper It - Microsoft Uncovers Sneaky New Attack Targeting Top LLMs to Gain Access to Encrypted Traffic

Why It Matters

Whisper Leak exposes a fundamental privacy weakness in AI chat platforms, enabling covert surveillance and eroding user trust, which could trigger regulatory scrutiny and force the industry to redesign streaming architectures. Mitigations will add latency and overhead, impacting performance and cost for providers.

Summary

Microsoft disclosed a new side‑channel attack dubbed “Whisper Leak” that can infer the topics of user conversations with large‑language models (LLMs) by analyzing the size and timing of encrypted packets sent during streamed responses. The technique exploits the incremental token transmission pattern of LLMs, allowing attackers—such as government agencies or ISPs—to make educated guesses about sensitive queries without breaking encryption. OpenAI, Mistral and xAI quickly rolled out mitigations that inject random-length text into replies to disrupt token‑size consistency, while Microsoft recommends VPN use, avoiding public Wi‑Fi, or opting for non‑streaming models. The finding highlights a broader privacy flaw in AI chat services that traditional security tools cannot detect.