paidf-auto-labeling
Pass
Audited by Gen Agent Trust Hub on Sep 21, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [SAFE]: The skill provides comprehensive operational guidance for NVIDIA's PAIDF Auto-Labeling system. It includes extensive guardrails, such as mandatory verification of environment prerequisites, confirmation of critical run inputs with the user, and the use of placeholders for sensitive credentials (e.g., NVIDIA_API_KEY). All examples follow security best practices by using local service URLs and environment variables for secrets.
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to process external media (video and images) and ingest data from specified paths, which creates a potential surface for indirect prompt injection if the processed content contains malicious instructions. However, this risk is mitigated by detailed prompt-authoring guidelines provided in the references. These guidelines instruct agents to follow evidence-first policies, utilize uncertainty handling, and avoid inferring intent or protected attributes from the data.
- Ingestion points: Data entries and media paths specified in YAML cookbooks (e.g., SKILL.md, references/stages/captioning.md).
- Boundary markers: Explicit evidence-first prompt structures and anti-fabrication rules defined in references/prompt-authoring/prompt-patterns.md.
- Capability inventory: Execution of containerized stages via the workflow-runner service using make run commands.
- Sanitization: Instructions for confirming critical inputs (data paths, endpoints) and prohibiting the inclusion of hardcoded secrets in configurations.
Audit Metadata