vlm-scene-captioning
Pass
Audited by Gen Agent Trust Hub on Sep 10, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes 3D scene data (USD files), specifically semantic labels and prim names, which are then used to construct prompts for an external LLM (NVIDIA NIM). Maliciously crafted labels within a USD file could potentially influence the LLM's output.
- Ingestion points: USD scene files (specifically semantic labels and prim names extracted from the stage).
- Boundary markers: No explicit delimiters or "ignore embedded instructions" warnings are described for the data ingested from the USD scene prims.
- Capability inventory: The skill uses the
CaptionAPIto perform network requests to an LLM provider and writes generated captions, scene graphs, and visualization images to the local file system. - Sanitization: No sanitization or filtering of prim labels or names is mentioned before they are interpolated into the prompt sent to the LLM.
Audit Metadata