command-injection

name	command-injection
description	SAST detection methodology for OS command injection (CWE-78), including shell metacharacter injection, argument/flag injection into git/curl/tar, wildcard injection, shell=True, and indirect injection via env vars and filenames. Use when reviewing code that spawns OS processes. Writes confirmed findings to findings/03-command-injection.md.

Command Injection — SAST Methodology

Class: OS Command Injection · CWE-78 · OWASP: A03 Injection Findings file: findings/03-command-injection.md

1. Overview

Command injection occurs when attacker-controlled data reaches a process-spawning sink and influences either the command string interpreted by a shell or the arguments passed to a program. Two distinct failure modes: (1) a shell parses the input and metacharacters (;, |, &, $(), backticks) start new commands; (2) even with no shell, input becomes an argument that the target program treats as an option (argument/flag injection). The core test: does a source reach an exec/system/spawn sink, and is a shell involved or can the input pose as a flag?

2. Where it lives

• Any process-spawn API: os.system, subprocess with shell=True, child_process.exec, Runtime.exec, ProcessBuilder, PHP system/exec/ passthru/shell_exec/backticks, Ruby system/backticks/%x, Go exec.Command with sh -c.
• Shell-outs to CLI tools (git, curl, tar, ffmpeg, imagemagick, ping, nmap, zip) with user-supplied args, filenames, or URLs.
• "Glue" scripts, webhooks, CI runners, and file-conversion/thumbnail pipelines.
• Template/report generators that call wkhtmltopdf, pandoc, latex, etc.

3. Sources (tainted input)

All HTTP inputs, plus indirect sources that are easy to miss: environment variables (multi-tenant/CI), filenames in uploads and archives (a file named ;rm -rf or --upload-file), image/document metadata read by a downstream tool, DB-stored values re-used to build a command (second-order), and hostnames/URLs passed to network tools. Trace stored/derived values back to their write source.

4. Sinks (dangerous operations)

# Python — dangerous
os.system("ping -c1 " + host)                     # shell parses metachars
subprocess.run(f"tar czf out.tgz {path}", shell=True)   # shell=True = shell
subprocess.call(["git", "clone", url])            # NO shell, but url="--upload-pack=..." = arg injection
os.popen("convert " + name + " out.png")

// Node — dangerous
child_process.exec(`ping -c1 ${host}`);           // exec spawns /bin/sh
child_process.execSync("tar czf o.tgz " + path);
spawn("sh", ["-c", `ls ${dir}`]);                 // explicit shell
spawn("git", ["clone", url]);                      // arg injection if url starts with -

// Java — dangerous
Runtime.getRuntime().exec("sh -c \"ping " + host + "\"");
new ProcessBuilder("bash", "-c", "grep " + q + " f").start();
Runtime.getRuntime().exec(new String[]{"tar","xf",userFile}); // arg injection

// PHP — dangerous
system("ping -c1 ".$_GET['host']);
exec("convert ".$file." out.png", $o);
passthru("git clone ".$url);
$out = shell_exec("ls ".$dir);
$x = `whois $domain`;                              // backticks

# Ruby — dangerous
system("ping -c1 #{host}")                         # interpolation + shell
`tar czf out.tgz #{path}`                          # backticks
%x(git clone #{url})

// Go — dangerous
exec.Command("sh", "-c", "ping "+host).Run()       // sh -c = shell
exec.Command("tar", "xf", userFile).Run()          // arg injection

5. Sanitizers / safe patterns

Safe:

• No shell + argument vector: pass a program and an array of args (subprocess.run([...], shell=False), execFile/spawn without a shell, ProcessBuilder(list), exec.Command(name, args...)). Metacharacters are inert.
• Use the -- end-of-options separator and/or validate that user args don't start with - to stop flag injection (git clone -- <url>).
• Strict allowlist for the command/subcommand; validate hostnames, paths, and IDs against a tight regex/enum.
• Prefer native libraries over shelling out (HTTP client vs curl, archive lib vs tar).

Fails / not a real sanitizer:

• Escaping quotes but still running through a shell — many metachars need no quotes ($(), `, \n, ;).
• Blocklisting ; and | while &&, \n, $(), ${IFS}, backticks slip through; blocklists are always incomplete.
• shlex.quote/escapeshellarg applied to one arg but the command still built as a shell string, or applied but the value is later re-split.
• Array/exec vector used, but the value is a flag — arg injection is unblocked by shell-escaping (there is no shell to escape).
• Validating the value but passing the raw filename from an upload/archive.

6. Detection methodology

1. Find sinks: grep the spawn APIs and shell indicators.

   rg -n 'os\.system|subprocess\.(run|call|Popen|check_output)|os\.popen'
   rg -n 'shell\s*=\s*True|child_process|\.exec(Sync)?\(|execFile|spawn\('
   rg -n 'Runtime\.getRuntime\(\)\.exec|ProcessBuilder|Process\.Start'
   rg -n 'shell_exec|passthru|\bsystem\s*\(|\bexec\s*\(|popen|proc_open'
   rg -n '`[^`]*\$\{|%x\(|\bsystem\s*\(|IO\.popen|Open3'
   rg -n 'exec\.Command\([^)]*"(sh|bash|cmd|powershell)"'
   

2. Determine if a shell is involved: shell=True, a string command, sh -c/cmd /c/bash -c, or an API that spawns a shell (exec, system, backticks). If yes → metacharacter injection is in scope.
3. If no shell, check argument injection: can input become a leading -/-- flag to the target tool? (git --upload-pack/--output, curl -o/--config, tar --to-command, find -exec, ssh -o).
4. Trace input to a source (direct, env, filename, or second-order). Stop at a constant.
5. Check for a real neutralizer: array vector and flag guard, or a strict allowlist. Verify it's on the reaching path and not bypassable (see §5).
6. Confirm reachability and note the platform (POSIX shell vs Windows cmd/PowerShell) — quoting and metachars differ.

7. Modern & niche variants

• Argument / flag injection (no metachars needed): input passed as a program argument that the program interprets as an option. git clone --upload-pack='sh -c ...' host:repo, curl --config /attacker or -o /path, tar --to-command=cmd, find . -exec, ssh -o ProxyCommand=, zip -T --unzip-command. Defeats shell-escaping entirely because there is no shell. Guard with -- and a leading-dash check.
• Wildcard injection: attacker controls filenames in a directory later globbed (tar czf a.tgz *, chown -R x *). A file named --checkpoint-action=exec=sh or -rf is expanded by the shell into an argument — classic tar/rsync/chown wildcard attacks.
• Indirect via environment variables: input written to an env var consumed by a child process (IFS, LD_PRELOAD, GIT_SSH, PAGER, PATH, PERL5OPT). The command string looks clean; the env is the vector.
• Indirect via filenames / archives: upload or zip-slip filename becomes an arg or is globbed. Metadata (EXIF, PDF fields) read by a shelled-out tool.
• shell=True and backtick/%x interpolation: the most common Python/Ruby/PHP root cause — a formatted string with a shell.
• exec vs execve semantics: Java Runtime.exec(String) naively splits on spaces (no shell) — quoting breaks, and exec(String[]) runs a program directly; neither parses metachars unless you explicitly invoke sh -c. Know which overload is used before judging.
• Windows cmd vs PowerShell: cmd /c metachars are &, |, %VAR%, ^; PowerShell adds ;, `, $(), |, and expression injection. .bat wrappers re-parse arguments (the "BatBadBut" class) even from array-vector spawns.

8. Common false positives

• Array/exec-file vector with no shell and no possibility of flag injection (value is not a leading arg, or --/dash-guard present).
• Command and all args are hard-coded constants.
• User input validated against a strict allowlist/enum before use.
• A native library is used and the "command" string is unrelated dead code.

9. Severity & exploitability

Base Critical — command injection is remote code execution. Confirm and it is Critical; pre-auth or default-config reachability keeps it there. Argument injection may be High if it yields file write/read or option abuse short of full RCE, rising to Critical when the flag reaches code execution (e.g. git --upload-pack). Lower only if reachable exclusively behind strong admin auth with constrained input. See references/severity-model.md.

10. Remediation

Avoid shelling out; use native libraries. When you must spawn, use the array/argument vector with shell=False and never build a shell string. Add a strict allowlist for commands and validate every argument; use -- and reject values starting with -. Never pass raw filenames/env into commands. Run with least privilege.

11. Output

Append each confirmed finding to findings/03-command-injection.md using references/finding-template.md. Set Class: OS Command Injection, CWE: CWE-78, and in Chain potential name primitives such as remote code execution (terminal primitive — usually the end of a chain), arbitrary file read/write (via flag injection / redirection), or note when it consumes a primitive like arbitrary file write or SSRF (a controlled URL fed to curl) to reach RCE.

Primitives (controlled): provides CODE_EXEC; consumes none

References

• CWE-78; OWASP A03:2021 Injection; OWASP Command Injection Prevention Cheat Sheet; GTFOBins/LOLBAS (argument-injection gadgets).