Private Insights into AI Use
summary
Describes a practical break of Anthropic's CLIO via the CLIOPATRA attack that seeds dozens of poisoned Claude chats to bypass PII scrubbing, clustering and LLM audits and, on synthetic medical conversations, extracts a target’s full medical history up to 100% extraction when the attacker knows basic demographics plus one symptom. Proposes Provably Private Insights (PPI): an open-source architecture combining TEEs for verifiable compute, “Data Expert” LLMs inside secure enclaves, and Differential Privacy, demonstrated at scale in Google’s Android Recorder app.
AI-written summary · about
chapters (8)
- 0:00AI insights in products
- 4:44Claude Insights system
- 9:06ClioPatra attack
- 16:12Extracting private data
- 24:50Differential privacy
- 27:24Confidential compute insights
- 30:17Private AI challenges
- 31:24Q&A
related talks
01h 20mUnlocking societal benefits of AI with privacy protective technologies.
43mPanel: Privacy Design Patterns for AI Systems: Threats and Protections
28mDon't leak user data to AI - Strategies for protecting PII from LLMs and MCP
24mAI, Health & Confidentiality
19mHow Can We Have Private Access to Intelligence?
42mPennywise - Invisible privacy risk in the new AI era
22mEnhancing Privacy through an Interactive On-demand Incremental Information
29mPersonal Data Pseudonymization Versus Anonymization In The Age Of AI & Big Data