The Agent Skills Directory

[DATA_EXFILTRATION]: The skill features an optional inference probe that sends the contents of analyzed files to an LLM endpoint (Ollama local instance or OpenAI-compatible cloud API) for analysis.
Evidence: In scripts/red.py, the _call_model function uses urllib.request.urlopen to POST file content to the configured llm_base_url when the --inference flag is active.
[COMMAND_EXECUTION]: The skill is implemented as a Python CLI script that performs recursive file system reads and network operations.
Evidence: The script uses argparse for CLI interactions and os/glob to collect files for processing.
[CREDENTIALS_UNSAFE]: The script provides standard support for using the OPENAI_API_KEY environment variable to authenticate cloud-based LLM requests.
Evidence: Line 159 in scripts/red.py retrieves the API key via os.environ.get("OPENAI_API_KEY", "") for inclusion in the request headers.
[SAFE]: The skill exhibits high security awareness, providing detailed instructions on its 'dual-use' nature and warning users to only run inference probes on synthetic or consented data. Its network activities are well-documented and consistent with the stated functionality.

red