We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
Hugging Face Blog·Jun 18·5 min read
MosaicLeaks: Can your research agent keep a secret?
TL;DR
Deep research agents risk leaking sensitive information through external queries, creating privacy concerns. MosaicLeaks introduces a new training method called Privacy-Aware Deep Research (PA-DR) to address this issue.
✦ Why It Matters
Engineers can implement PA-DR to enhance privacy in AI systems handling sensitive data.
Key Takeaways
- Integrate PA-DR to balance task performance and privacy.
- Measure leakage using intent, answer, and full-information categories.
- Avoid training solely for task performance to prevent increased leakage.
- Utilize situational task rewards for more efficient training.
How It Works
MosaicLeaks evaluates privacy risks by analyzing how agents construct web queries over time. It uses a reward system that penalizes queries leading to potential leakage, allowing agents to learn safer query construction while maintaining performance.