The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands using subprocess.run with a command string constructed via f-string interpolation. Specifically, the model parameter, which is a user-provided command-line argument, is directly embedded into a bash -c command without sanitization. This allows an attacker to inject arbitrary shell commands (e.g., by providing a model name like 'dummy; malicious_command;').
Evidence: In run.sh, the embedded Python script contains the run_live function which executes result = subprocess.run(["bash", "-c", f"... --model {model} ..."], ...).

benchmark-models