Dev.to — MCP Server Audit Finds 66% Have Critical Vulnerabilities

AI relevance: MCP servers provide the tool layer that LLM agents use to interact with systems — security flaws in these servers translate directly to agent compromise (RCE, data exfiltration, credential theft).

  • Yedan Yagami built and deployed 9 MCP servers on Cloudflare Workers, then audited them using the OWASP Agentic AI Top 10 framework.
  • The audit found that 66% of MCP servers have critical vulnerabilities, with shell/command injection being the most prevalent at 43%.
  • 30 CVEs were filed against MCP servers in January-February 2026 alone, indicating rapid growth of security issues in the ecosystem.
  • The average security score was 34/100 across all audited servers.
  • Key vulnerability categories found:
    • Shell/Command Injection (43%) — user input passed directly to shell commands
    • Tooling Infrastructure Gaps (20%) — missing security controls
    • Authentication Bypass (13%) — weak or missing authentication
    • Path Traversal (10%) — insufficient path validation
    • Prompt Injection via Tools (8%) — malicious instructions in tool descriptions
    • Data Exfiltration (6%) — unrestricted network access
  • A particularly concerning finding: 84.2% of tool poisoning attacks succeed when auto-approval is enabled (per Trend Micro research).
  • The author developed a 30-item security checklist covering all vulnerability categories with severity ratings and remediation code examples.

Why it matters

  • MCP has exploded with 97 million monthly SDK downloads and adoption by every major AI company — making server security critical infrastructure.
  • Shell injection vulnerabilities mean an AI agent could be tricked into executing arbitrary commands like rm -rf / through seemingly benign tool calls.
  • Prompt injection via tool descriptions allows malicious servers to hijack agent behavior by embedding instructions in tool metadata.
  • The lack of standard security scanning infrastructure for MCP servers creates a supply chain risk for AI agent deployments.

What to do

  • Implement input sanitization: Use parameterized execution instead of string concatenation for shell commands.
  • Scan tool descriptions: Check for prompt injection patterns in MCP tool metadata before deployment.
  • Use OAuth 2.1 + PKCE: Follow MCP spec recommendations for secure authentication.
  • Restrict network egress: Whitelist domains that MCP tools can connect to prevent data exfiltration.
  • Bind to localhost only: Avoid exposing MCP servers to external networks unless absolutely necessary.
  • Audit existing servers: Use the 30-item checklist to evaluate your current MCP deployments.

Sources