i4h-workflow-dataset-annotate

Pass

Audited by Gen Agent Trust Hub on Sep 9, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill clones the i4h-workflows repository from https://github.com/isaac-for-healthcare/i4h-workflows. This is a vendor-owned resource associated with the Isaac for Healthcare project.
  • [COMMAND_EXECUTION]: The skill executes orchestration and utility scripts (vllm.sh and run.sh) located within the cloned repository to manage VLM serving and annotation tasks.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes robotic recordings (HDF5) and environment configurations (YAML), passing task descriptions to a VLM for success classification. This represents a potential vulnerability surface if the processed data contains malicious instructions.
  • Ingestion points: Reads task success criteria from config/environments/<env>.yaml and frame data from HDF5 files specified by the HDF5_PATH variable.
  • Boundary markers: None identified; instructions do not explicitly define delimiters for external content within the VLM prompt.
  • Capability inventory: Performs network requests to a VLM API endpoint (VLM_BASE_URL), writes annotation results and logs to local storage, and executes shell scripts.
  • Sanitization: The skill does not document explicit sanitization or filtering of the task descriptions or HDF5 metadata before they are interpolated into the VLM workflow.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 9, 2026, 07:47 AM
Security Audit — agent-trust-hub — i4h-workflow-dataset-annotate