DEV Community

Aurum Revenue
Aurum Revenue

Posted on

Your subprocess call is a shell injection waiting to happen

A few weeks ago a build script of mine died with a bash syntax error. The command looked harmless:

prompt = "compute (17*23) and print it"
subprocess.run(f"echo {prompt}", shell=True)
Enter fullscreen mode Exit fullscreen mode

Bash saw (17*23) and tried to parse a subshell. The script aborted. Worse: if that string had come from a user, `rm -rf ~` would have executed.

This is the single most common way Python code accidentally becomes an attack surface. shell=True plus an f-string is a remote code execution primitive dressed up as convenience.

The usual advice, and why it is not enough

"Use a list, never shell=True." Correct, and still incomplete:

subprocess.run(["echo", "(17*23)"])   # safe
subprocess.run(["sh", "-c", user_input])  # back to square one
Enter fullscreen mode Exit fullscreen mode

The moment you legitimately need a shell — pipes, globs, a tool that only speaks shell — you are back to building a string. And building a shell string by hand is where people get it wrong.

What actually goes wrong

MSYS/git-bash on Windows expands * and ~ in argv before your program sees them. shlex.quote protects against the shell, but not against a layer that rewrites arguments before the shell runs. Quoting rules differ between cmd.exe, PowerShell, and POSIX sh. A single ! can trigger history expansion in an interactive shell.

I wrote shellguard to make this boring:

import shellguard as sg

sg.run(["echo", "(17*23)"])          # -> '(17*23)\n'
sg.run(["echo", "`echo INJECTED`"])  # -> literal text, nothing executed
Enter fullscreen mode Exit fullscreen mode

Zero dependencies. It builds the command line correctly for the shell you are actually targeting, and it refuses to silently interpolate untrusted data into a shell string.

The part I care about most

The test suite is 26 real attack cases. Not mocked. Each test spawns an actual subprocess and asserts on the real exit code and real stdout:

  • command substitution with backticks and $()
  • chained commands with ;, &&, ||
  • glob and tilde expansion under MSYS
  • newline and carriage-return injection
  • quote-breaking with mixed ' and "

If a payload executes, the test fails. That is the only assertion that matters for a security library.

Try it

If you have ever written f"grep {user_input} file.txt", this is for you.

What is the worst shell-quoting bug you have shipped? I would genuinely like to know — the test suite grows from real incidents.

Top comments (0)