The Agent Skills Directory

[COMMAND_EXECUTION]: The Python script model_promotion_gate.py utilizes subprocess.run to execute security tools including cosign and trivy. The implementation uses list-based arguments rather than shell strings, which is a recommended practice to prevent command injection.
[INDIRECT_PROMPT_INJECTION]: The skill describes workflows that ingest untrusted external data such as model weights, configurations, and attestations.
Ingestion points: model_promotion_gate.py (accepts image identifiers via CLI), fickling (processes pickle files), and cosign (verifies attestations).
Boundary markers: None explicitly defined in the provided instructional code.
Capability inventory: The skill utilizes subprocess execution for tool orchestration, network operations for OCI registry interactions, and file system access for artifact signing.
Sanitization: While the Python script uses safe subprocess patterns, the promotion logic primarily validates the existence and return codes of security checks rather than performing deep schema validation of the ingested metadata payloads.
[EXTERNAL_DOWNLOADS]: The documentation suggests installing the fickling security package via pip and references standard GitHub Actions (actions/checkout@v4) for CI/CD workflows.

model-supply-chain-security