Capture Notes

The paper introduces MCPPrivacyDetector, a context-aware static analysis framework for privacy leakage risks in MCP servers.

AI security relevance:

• Frames MCP leakage as protocol-induced: sensitive values may cross the local/LLM boundary through tool returns, logs, or exceptions even without explicit outbound network exfiltration.
• Reports analysis over 10,655 real-world MCP servers and leakage rates above 10%.
• Important for MCP server vetting, tool handler taint analysis, and agent privacy boundaries.

Suggested ingest focus:

• Evidence for MCP as a local-to-global data boundary.
• Add to MCP security, agent tool security, and non-human identity/credential exposure topics.