tao-generate-image-grounding
Pass
Audited by Gen Agent Trust Hub on Sep 23, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-supplied image captions and image paths via a VLM, creating a surface for indirect prompt injection where malicious instructions could be embedded in the data. \n
- Ingestion points: External data enters the agent context through the
data.input_jsonlfile (SKILL.md). \n - Boundary markers: The skill mentions caption cleaning in the first step, but no explicit prompt delimiters or "ignore embedded instructions" markers are provided for the VLM prompt. \n
- Capability inventory: The skill executes the
auto_labelCLI tool and performs file system writes for results and checkpoints. \n - Sanitization: The first pipeline step normalizes captions to extract referring expressions. \n- [EXTERNAL_DOWNLOADS]: The skill references and downloads official NVIDIA container images from the
nvcr.ioregistry and instructions for installing thevLLMpackage. These originate from trusted vendor and well-known service resources. \n- [COMMAND_EXECUTION]: The skill usesbashto execute Docker commands and theauto_labeltool to run the image grounding pipeline and manage inference services.
Audit Metadata