Security scan results for typescript-sdk — MCPSafe AIVSS 53/100 (Grade F)

[mcpsafe-gh]

Hi team 👋

I ran a free deep security scan of modelcontextprotocol/typescript-sdk using MCPSafe — a purpose-built scanner for MCP servers using a 5-LLM consensus panel to detect prompt injection risks, over-scoped tool schemas, supply chain issues, and more.

Results: 53/100 · Grade F

Severity	Count
🔴 Critical	0
🟠 High	11
🟡 Medium	125
🟢 Low	0

Summary: 11 high + 125 medium findings across the SDK — worth reviewing before downstream MCP servers adopt these patterns

📋 Full report with findings and evidence: https://mcpsafe.io/registry/github/modelcontextprotocol/typescript-sdk

---

Add a security badge to your README

[![MCPSafe](https://api.mcpsafe.io/badge/github/modelcontextprotocol/typescript-sdk.svg)](https://mcpsafe.io/registry/github/modelcontextprotocol/typescript-sdk)

This badge auto-updates whenever a new scan runs — great for showing users and enterprise customers your security posture at a glance.

---

Feel free to close this if you're already tracking these findings. Happy to answer any questions about specific findings.

— Truong BUI · mcpsafe.io

[redbotster]

The 11 high findings are worth digging into. A few patterns that tend to produce clusters of findings in MCP SDKs at this stage: tool schemas that over-expose parameter descriptions (prompt injection surface in the schema itself), token handling without scope validation at the tool level, and credential pass-through that assumes transport security is sufficient for secrets at rest.

The medium-count being 10x the high count is a typical shape for infrastructure-layer SDKs — lots of defense-in-depth gaps that aren't individually critical but compound when stacked. Downstream MCP servers building on these patterns inherit them.

We've been building specifically on this surface at 1Claw (Vault for credential brokering, Shroud for the inspection layer). Happy to compare notes on the high findings if it's useful. docs.1claw.xyz