nvidia-ontology-management
Pass
Audited by Gen Agent Trust Hub on Oct 10, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill defines the semantic layer that governs agent reasoning and data interpretation, creating an inherent surface for indirect prompt injection where malicious instructions could be embedded in ontology definitions.\n
- Ingestion points: External, potentially untrusted data enters the context via YAML model imports (
/api/model/import), SQL attribute expressions (/api/sql-attributes), and LLM-generated suggestions for attribute descriptions (/api/sql-attributes/{id}/description-suggestion) as detailed inreferences/write-api.md.\n - Boundary markers: The instructions in
SKILL.mdrequire the agent to disclose proposed changes to the user for review, andreferences/publication.mdmandates a staged exact readback comparison to verify that changes persisted correctly without unintended side effects.\n - Capability inventory: The skill can perform full CRUD operations on the semantic layer (Terms, SQL Attributes, Column Attributes) and can trigger asynchronous, destructive global compilation resets via
/api/semantic-compilation/reset.\n - Sanitization: The skill implements a mandatory validation step using the
/api/sql-attributes/validateendpoint to check SQL expressions for parse failures and validity before they are applied to the ontology.
Audit Metadata