ACB-MCP-002: MCP server launch commands are not interpreted by a shell
| Field | Value |
|---|---|
| Status | draft |
| Severity | high |
| Check type | static |
| Applies to | mcp-server-declaration |
| Profiles | strict-general |
| Exceptions | permitted |
Description
A local MCP server is launched by naming an executable in command and passing its arguments as a list. The command is not a shell or command interpreter given a script string, and it contains no shell operators.
Rationale
A shell invocation turns every argument into code. Anything that reaches the string, such as an interpolated variable or a value an attacker can influence in a shared configuration file, runs with the user's privileges when the client starts the server. Passing arguments as a list to a named executable removes that interpretation step and makes the launched program visible in review.
Audit procedures
no-shell-interpretation (structure)
For each server entry with a command key, the command is not a shell or command interpreter (sh, bash, zsh, dash, ash, ksh, fish, csh, tcsh, cmd, powershell, pwsh) invoked with a string-evaluation flag (-c, /c, /k, -Command, -EncodedCommand), whether directly or through a wrapper such as env or sudo, and the command value contains none of the shell operators ; & | < > $ ` or a newline.
Illustrative failing indicators (non-exhaustive):
"command": "sh""command": "env""-c""command": "node server.js && curl https://example.com"
Remediation
Name the server executable directly in command and move each argument to its own element of args. If a wrapper script is needed, ship it as a file in the repository, review it as code, and launch it by path.
References
- OWASP Top 10 for Agentic Applications 2026: ASI05
- OWASP MCP Top 10: MCP05
- Careful Adoption of Agentic AI Services risk areas: design-and-configuration
- NIST CSF 2.0: PR.PS-01, PR.PS-05
- https://www.nsa.gov/Portals/75/documents/Cybersecurity/CSI_MCP_SECURITY.pdf