The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/init-dvc-project.sh is vulnerable to shell command injection through the --remote argument. It uses an internal run function that calls eval "$@" on a string containing the $REMOTE_URL variable, allowing execution of arbitrary commands if the URL contains shell metacharacters like semicolons or backticks.
[COMMAND_EXECUTION]: The script scripts/queue-helper.sh contains a command injection vulnerability in the enqueue and grid subcommands. The script constructs a $flags variable from user-provided parameters and executes dvc exp run --queue $flags without quotes. This allows an attacker to inject shell commands by providing a parameter value that includes shell operators (e.g., model.lr=1; rm -rf /).

dvc-ml-workflow