VulnerableMCP — MCP security database for real-world tool flaws

AI relevance: MCP servers are the tool layer for LLM agents, so a centralized vulnerability database directly maps to real-world risk in agent deployments.

  • VulnerableMCP publishes a rolling catalog of MCP-related issues with impact, exploitability, and prevalence tags.
  • The database pulls in entries from CVE feeds, advisories, and security research, not just vendor blogs.
  • Recent examples listed include cross-client response leakage in the MCP TypeScript SDK and unauthenticated RCE via exposed MCP servers.
  • Each entry calls out attack class (prompt injection, command injection, SSRF, DNS rebinding) to help triage by threat model.
  • Entries explicitly note data exfiltration, credential theft, or tool abuse impacts where relevant.
  • The project acts as a shared memory for MCP supply-chain and tool-layer risks across vendors.

Why it matters

  • Agent platforms often reuse MCP servers across teams; a single vulnerable tool can become a fleet-wide blind spot.
  • Security teams need a single source of truth to map agent tooling risks to concrete mitigations.
  • Consolidated issue tracking improves patch prioritization and reduces reliance on scattered blog posts.

What to do

  • Inventory: Compare your installed MCP servers/skills against VulnerableMCP’s database.
  • Gate installs: Require review for MCP packages with command execution, network fetch, or filesystem access.
  • Harden defaults: Disable 0.0.0.0 bindings, enforce auth, and restrict tool permissions by default.

Sources