About this tag
The tag 'llm privacy' covers discussions about privacy vulnerabilities in large language models, particularly a disclosed attack called Whisper Leak. This attack exploits encrypted streaming traffic between users and remote LLMs to infer the topic of a user's prompt without accessing plaintext content. The content focuses on how Microsoft researchers identified this class of vulnerability, highlighting risks in cloud-hosted LLM services that use streaming APIs. Discussions center on the metadata leakage from encrypted traffic and its implications for user privacy. The tag is relevant for users concerned about security and privacy when using AI-powered language models, especially in enterprise or personal contexts where sensitive topics may be discussed.
  1. WindowsForum AI

    LLM Privacy Policies: 13 of 20 Services Train on User Inputs

    Cybersecurity firm Bridewell studied the privacy policies of 20 popular large language model (LLM) services and found that they average 4,603 words. Reading one takes just under 20 minutes, and the average readability score is 40.2, a level Bridewell describes as best suited to university...
  2. WindowsForum AI

    Whisper Leak: Encrypted LLM Traffic Reveals Topic Metadata

    Microsoft researchers have disclosed a new class of privacy vulnerability — dubbed Whisper Leak — that turns encrypted streaming traffic between users and remote large language models (LLMs) into a surprisingly effective intelligence source for eavesdroppers, enabling an adversary to infer the...