Don't leak user data to AI - Strategies for protecting PII from LLMs and MCP
summary
Live demo shows an LLM using MCP (Model Context Protocol) listing schema objects, pulling a customer row and leaking PII into Claude servers; removing table visibility prevents leakage but triggers hallucinations and broken queries. Recommended controls are don't touch prod: run LLMs against non‑prod instances, use restricted connection strings, or expose anonymized views generated with PG Anonymizer v2 (Rust) or Faker-backed masks. Options compared: static irreversible masking, dynamic role‑based masking, and dump‑anonymize‑restore; logical replication masking is coming.
AI-written summary · about
chapters (14)
- 0:00Introduction and speaker welcome
- 1:35Personal identifiable information risks
- 4:00EU AI Act and developer responsibilities
- 7:12Avoiding production database access
- 8:14Model Context Protocol overview
- 10:21Vibe coding demo with PostgreSQL MCP
- 13:59PII leakage and its consequences
- 14:41Restricting LLM database access
- 18:38PostgreSQL Anonymizer introduction
- 21:28Random views for safe data sharing
- 23:11Dynamic masking approach
- 24:39Tips for reliable AI integration
- 25:41Long-term anonymization strategies
- 26:23Q&A session
related talks
28mAnonymization in the LLM Era: No Room for Compromise
06mData Privacy Concerns and Solutions When using LLMs in Production
12mStriking the Balance: Safeguarding Customer Privacy While Empowering Employees
15mHarnessing LLMs for Scalable Data Minimization
32mPrivate Insights into AI Use
28mOpen the gate a little: strategies to protect and share data
44mI Abuse Your LLM Live with 10 Security Vulnerabilities
50mSecuring LLM-Powered Applications: Overcoming Security and Privacy Challenges