The Agent Skills Directory

[COMMAND_EXECUTION]: Several scripts (runner.py, classifier.py, scenario_generator.py, spec_generator.py) use subprocess.run to execute external tools like claude, git, and setup utilities. While runner.py implements an ALLOWED_SETUP_EXECUTABLES whitelist for scenario setup, the primary execution path involves running shell commands based on data derived from LLM outputs.
[DYNAMIC_EXECUTION]: The skill uses LLMs to dynamically generate compliance specifications and test scenarios in YAML format. These generated specifications directly control the agent's behavior and the commands executed in the sandbox environment. Although yaml.safe_load is used, the logic flow relies on the integrity of the LLM-generated instructions.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process arbitrary .md files (skills/rules) to generate tests. If a target file contains adversarial instructions, it could influence the spec_generator.py or scenario_generator.py to produce malicious test cases or setup commands that attempt to bypass the sandbox restrictions.

skill-comply