Whisper Leak Reveals AI Conversation Topics Remotely
Microsoft researchers analyzed token-stream timing from 30 AI models using 11,800 prompts and identified a 'Whisper Leak' side-channel that can reveal conversation topics. They report detection accuracies for dangerous queries ranging 71–100%, with 19 models exceeding 97%, and successful detection in noisy 10,000-conversation tests for several models. Providers are deploying packet-padding mitigations; users can switch to local models, disable streaming, or use VPNs.
Key Points
- 1Demonstrates timing-based side-channel leaks by analyzing token-stream packet patterns across 30 models and 11,800 prompts
- 2Shows high topic-detection accuracy (71–100%), with 19 models exceeding 97% for dangerous queries
- 3Implies operators must pad or disable streaming; users should prefer local models or VPNs for sensitive chats
Scoring Rationale
Comprehensive Microsoft study reveals systematic side-channel risks with practical mitigations; effectiveness varies by model and server configuration.
Sources
Public references used for this report.
Practice with real Streaming & Media data
90 SQL & Python problems · 15 industry datasets
250 free problems · No credit card
See all Streaming & Media problems
