nvidia-ontology-query
Pass
Audited by Gen Agent Trust Hub on Nov 4, 2026
Risk Level: SAFEDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: The skill facilitates network communication with Auto Ontology API endpoints (e.g.,
ontology.example.com/api/chat/completions). It uses standard authentication mechanisms, including environment variables (e.g.,AUTO_ONTOLOGY_API_TOKEN) and HTTP headers (Authorization: Bearer,x-api-key). These operations are consistent with the skill's primary purpose of providing remote data access. - [INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted data from database query results. To mitigate risks such as data fanout or misleading measures, it implements a comprehensive validation framework in
references/query-validation.md. - Ingestion points: Data enters the context via SQL query results (
rows) and SSE streams from the Auto Ontology API. - Boundary markers: The instructions require the agent to restate material constraints (population, grain, unit) and check them against the SQL before presenting answers.
- Capability inventory: The skill uses
requestsfor network operations. No direct file-system writes or arbitrary command execution capabilities are present. - Sanitization: The skill relies on manual and LLM-driven validation steps (comparing SQL to intended constraints) to detect data poisoning or logic errors.
- [CREDENTIALS_UNSAFE]: The documentation in
references/stack.mdprovides architectural details regarding the authentication contract, noting that the audience (aud) claim is not checked for specific SSO bearer tokens forwarded by internal NVIDIA services. This is a vendor-documented design specification for their SSO integration and does not constitute a vulnerability in the skill's implementation.
Audit Metadata