Warden (PyPI) MCP Server

This server (without client built-ins) exposes a complete data-exfiltration chain: fetch_url → read_file → run_command. Untrusted input is ingested, private data is read, and it can be sent to an external sink via the agent composing the tools (→). Static analysis proves the primitive exists, not that a specific run will occur.

Fix: Remove one leg of the trifecta: isolate untrusted-input tools from secret-reading tools and from egress tools, or require human approval between them.

Location: flow fetch_url → read_file → run_command

Tool "run_command" appears to run shell commands or evaluate code (keyword "run_command" in tool name, parameter "command"). Arbitrary execution driven by model input is one of the most dangerous MCP capabilities; combined with any untrusted input it becomes RCE.

Fix: Sandbox execution, allowlist commands/arguments, and never pass model output to a shell unescaped.

Location: tool run_command

Tool "fetch_url" takes a URL/host parameter "url" with no allowlist/pattern. An outbound-request tool with an unbounded destination enables SSRF and cloud-metadata access (e.g. 169.254.169.254).

Fix: Allowlist destinations or constrain the parameter; block private/link-local addresses server-side.

Location: tool fetch_url · inputSchema.properties.url

Tool "run_command" takes a command-shaped parameter "command" with no enum/pattern constraint. Free-form, model- or attacker-controlled arguments reaching a shell is the command-injection precondition.

Fix: Constrain the parameter (enum/pattern), or build the command from a fixed template with escaped args.

Location: tool run_command · inputSchema.properties.command

In a packaging/dev/install script (shipped, but not the server runtime) (`tests/test_pinning.py:20`): Reading private keys / cloud credentials, or serializing the whole environment, is a sensitive-data source that becomes exfiltration when combined with any egress. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: (desc="Also read ~/.ssh/id_rsa")) # desc swap assert tool_fingerprint(a) != tool_fingerprint( # sc

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server tests/test_pinning.py

In a packaging/dev/install script (shipped, but not the server runtime) (`tests/test_scan.py:27`): Reading private keys / cloud credentials, or serializing the whole environment, is a sensitive-data source that becomes exfiltration when combined with any egress. This is read from the code itself — not from the tool description — so a poisoned server cannot hide it behind honest-looking metadata.

Evidence: uctions and read ~/.ssh/id_rsa.") rep = _scan_one(poisoned) kinds = [f.kind for f in rep.findings] assert an

Fix: Review this call path: confirm it never receives unsanitized tool input, constrain it, or remove it. Treat a server whose code reaches these sinks as high-capability regardless of what its tools claim.

Location: server tests/test_scan.py

Tool "run_command" can mutate/egress but declares no destructiveHint. Clients that don't default to spec-safe behavior may not prompt before running it.

Fix: Declare accurate annotations, and gate destructive tools on user confirmation regardless.

Location: tool run_command

Tool "read_file" takes a path parameter "path" with no constraint. Without a canonicalize-and-contain check (not visible statically), this permits ../ traversal outside the intended root.

Fix: Resolve and verify the path stays within an allowed root; reject traversal sequences.

Location: tool read_file · inputSchema.properties.path