Fix: flip the default to SecurityConfig{Enabled: true}. Secure mode should be the operating default — not an env var users have to know to set. The --allow-unsafe flag (or equivalent env var) can preserve the unrestricted mode for developers who explicitly accept the risk, but that should require affirmative opt-in, not silence.
The second issue affects Docker users who do follow the security.yaml example. The official security.yaml — baked into the Docker image via COPY security.yaml /etc/mcp-shell/security.yaml — includes both /bin/bash and /usr/bin/python3 in allowed_executables. In secure mode (use_shell_execution: false), executor.go lines 142–163 parse the command and exec it directly:
The parseCommand() function uses strings.Fields() — split on whitespace — and the metacharacter check in containsDangerousShellConstructs() blocks |, &, ;, $, and similar constructs. With /bin/bash in the allowlist, the following call:
shell_exec(command="/bin/bash -i")
Parses to executable="/bin/bash", args=["-i"]. The executable is on the allowlist. -i contains no blocked metacharacters. The call passes all validation and executes as:
cmd = exec.CommandContext(ctx, "/bin/bash", "-i")
That's an interactive bash shell — stdin is shared with the mcp-shell process, which is the MCP command channel. The LLM now has a direct read/write channel to bash. The Python path is equally direct: shell_exec(command="/usr/bin/python3 /workspace/payload.py") where the payload file was written in a prior tool call. Both bypass all of secure mode's metacharacter filtering because the interpreter absorbs the dangerous content, not the direct command string.
The Docker image ships this config as the default. Any operator who runs the official image without a custom security.yaml is running with /bin/bash and /usr/bin/python3 in their allowlist — the advertised secure mode is not providing the restriction it claims.
Fix: remove /bin/bash, /bin/sh, and /usr/bin/python3 from allowed_executables in the default security.yaml. Shell interpreters defeat executable-allowlisting by design — the interpreter executes whatever it's handed, so allowing it is equivalent to disabling the allowlist entirely. The default config should contain only narrow utility binaries that can't themselves spawn arbitrary processes (ls, cat, grep, head, wc, date, pwd). A comment in the example config is also warranted:
# WARNING: Never add shell interpreters (bash, sh, python, perl, ruby) to
# allowed_executables. Doing so bypasses secure mode — the interpreter
# executes arbitrary code regardless of what metacharacter checks see.
Taken together: the from-source path runs with no security config by default, and the Docker path runs with a security config that includes shell interpreters. The two failure modes cover the full install surface. The maintainer clearly cares about security — the test coverage for injection scenarios and the containsDangerousShellConstructs implementation show the right intent. These are fixable with a one-line default flip and a config edit.
Addendum 2026-05-03 — counter to the sandbox-mitigation framing
Why the sandbox recommendation does not close either of the two findings above, with inline the PoC that demonstrates it.
The sandbox-mitigation framing assumes the operator (a) knows they need to sandbox, (b) sets up the sandbox correctly, and (c) the sandbox boundary actually contains what an attacker can reach via the MCP tool. None of those hold for the documented install path:
Default-off applies before the sandbox question is even asked. The from-source path (git clone && make install && mcp-shell) runs with Enabled: false from config.go:49. The MCP client config example in the README passes only MCP_SHELL_LOG_LEVEL, no MCP_SHELL_SEC_CONFIG_FILE. An operator following the documented install runs an unrestricted shell — the sandbox recommendation lives in a separate document the operator hasn't read yet because they've followed the README and it works. This is the from-source case in Finding 1 and it bypasses any recommendation that lives outside the install path.
Bash-in-allowlist defeats secure mode regardless of the sandbox. Even if the operator runs inside a Docker/firejail/bubblewrap sandbox, the official security.yaml includes /bin/bash and /usr/bin/python3 in allowed_executables. Inside that sandbox, the LLM can issue shell_exec(command="/bin/bash -c '<arbitrary>'") — passes allowlist, executes verbatim, and the sandbox boundary is the outer containment, not the intra-sandbox containment. The LLM still has full read/write to whatever lives inside the sandbox: secrets mounted into the container, environment variables, network access from inside the sandbox, files in the working directory. "Sandbox" only contains the blast radius if the threat model is "RCE escapes the host," which is not the threat model here — the threat model is "unintended command execution from the LLM session," which happens inside the sandbox boundary.
PoC fired locally — non-destructive marker write, mirrors the validateCommand → executeSecureCommand chain in Python (Go subprocess semantics for exec.CommandContext are equivalent to Python subprocess.run for arg-array dispatch):
The cmd-unfurl/expansion approach you raised is more interesting on the technical merits — it would close the bash-allowlist case directly (unfurl /bin/bash -c '<inner>' to expose the inner command for blocklist evaluation). It still wouldn't close the default-off case, because unfurl only runs when validation runs, and validation short-circuits when Enabled=false.
The minimal-change fix on both fronts remains: flip Enabled to true by default, drop shell interpreters from the example allowlist. Sandbox recommendation is reasonable as defense-in-depth but doesn't substitute for closing the two install-path defaults.