tao-generate-referring-expressions
Pass
Audited by Gen Agent Trust Hub on Jun 23, 2026
Risk Level: SAFE
Full Analysis
- [COMMAND_EXECUTION]: The skill executes the
auto_labelcommand-line utility (part of the NVIDIA TAO Toolkit) to perform the annotation generation. This is the primary intended function of the skill. - [EXTERNAL_DOWNLOADS]: The documentation in
references/vllm_server.mdprovides instructions for downloading official Docker images (vllm/vllm-openai) and Python packages (vllm) from well-known registries. These are standard dependencies for hosting self-hosted inference servers. - [CREDENTIALS_UNSAFE]: The skill requires API keys for VLM access (e.g.,
GOOGLE_API_KEYor OpenAI keys). The instructions correctly recommend using environment variables for secret management, though they also allow for configuration via YAML files, which is a common pattern for local development specifications. - [DATA_EXFILTRATION]: The skill transmits image data and bounding box labels to configured VLM endpoints (such as Google Gemini or OpenAI-compatible services). This is the documented and necessary behavior of the pipeline to generate the requested descriptions and captions.
- [INDIRECT_PROMPT_INJECTION]: The skill has a data ingestion surface as it processes external image files and KITTI-format label files. While these inputs could theoretically contain instructions designed to influence the VLM's output, this represents a low-severity inherent risk of VLM-based processing rather than a flaw in the skill's design.
Audit Metadata