The Agent Skills Directory

[COMMAND_EXECUTION]: Dynamic execution of shell commands in scripts/runner.py\n
The function _setup_sandbox iterates through setup_commands and executes them using subprocess.run after parsing with shlex.split.\n
These commands are generated by an LLM in scripts/scenario_generator.py based on the content of the skill or rule file provided by the user.\n
This pattern is vulnerable to indirect command injection, as a malicious input file could influence the LLM to generate harmful shell commands (e.g., file deletion or data exfiltration) that are then executed on the host system.\n- [COMMAND_EXECUTION]: Systematic use of subprocess.run to invoke external CLI tools.\n
Several scripts (runner.py, classifier.py, spec_generator.py, scenario_generator.py) execute the claude CLI to process prompts and generate outputs.\n
The runner.py script also executes git init within the sandbox directory.\n- [DATA_EXFILTRATION]: Processing of local configuration files.\n
The skill's primary workflow involves reading user-defined rules and agent settings (e.g., from ~/.claude/rules/) to evaluate compliance.\n
This involves ingesting potentially sensitive local information into the LLM context to generate scenarios and specifications.

skill-comply