Episode Details

Model Context Protocol servers are multiplying fast — 97 million monthly SDK downloads, over 10,000 public servers — but the protocol assumes every server is trustworthy. Tool descriptions load into the model's reasoning context with system-prompt authority, and there's no cryptographic attestation or native injection detection to catch tampering. This episode works through what MCP-specific observability actually requires: tool-call logging rather than HTTP logging, agent identity as a first-class principal, client-side hooks for the 86% of servers running locally, and write-time signed audit trails. Then the attacks that break conventional monitoring — tool poisoning, rug pulls like CVE-2025-54136, tool shadowing, and the Microsoft-described pattern where every individual action is authorized but the combination isn't. Plus the exposure numbers: 492 unauthenticated MCP servers found exposed to the open internet, a count that nearly tripled on rescan. Anchored by three products with three different postures — Mintive, Obot, and Docker MCP Gateway. Key sources: • MCP spec, Security and Trust & Safety section — https://modelcontextprotocol.io/specification/2025-06-18 • Mintive SIEM export docs (OTLP/Splunk HEC schemas) — https://docs.mintmcp.com/docs/siem-export • Mint Guard docs — https://www.mintmcp.com/docs/mint-guard Episode 5973 — myweirdprompts.com/n/5973 (permanent short link: myweirdprompts.com/403507)