The Agent Skills Directory

[SAFE]: The core logic of the skill is contained in the check_model_provenance.py script, which is implemented using only Python standard library modules (argparse, json, re, sys, pathlib). It does not require external package installations or remote script execution.
[SAFE]: The skill is designed to be "network-free." It performs no HTTP requests or network operations, preventing the exfiltration of environment data or the resolution of licenses from external sources.
[SAFE]: Instructions within SKILL.md and skill.yml explicitly define safety boundaries, forbidding the downloading or execution of the third-party models being audited. This prevents the agent from being lured into running malicious binary artifacts.
[INDIRECT_PROMPT_INJECTION]: The skill handles untrusted data by reading a JSON model dossier provided by the user.
Ingestion points: The check_model_provenance.py script ingests data via a JSON file path passed to the --dossier argument.
Boundary markers: Not applicable as the input is structured JSON data rather than natural language prompts.
Capability inventory: The skill uses Read, Write, Edit, and Bash tools to manage the dossier and run the audit script.
Sanitization: The script uses standard json.loads for parsing and applies deterministic token-based matching for dataset names. The agent is instructed to treat the dossier as a data record for the audit script, and the script's logic does not interpret the content as instructions for the agent, minimizing the risk of indirect injection triggering unauthorized actions.

model-sourcing