i4h-workflow-dataset-annotate
Pass
Audited by Gen Agent Trust Hub on Sep 30, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill clones the i4h-workflows repository from GitHub to establish the necessary toolset for annotation. This repository is controlled by the project maintainers associated with the vendor.\n- [COMMAND_EXECUTION]: Utilizes shell commands for repository management (git) and executes project-specific Python tools (i4h-annotator, i4h-dataset) using the uv toolchain.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes external workflow recording data and user-provided success criteria through a VLM, creating a potential vector for indirect instructions.\n
- Ingestion points: HDF5 recording episodes containing raw camera frames and user-provided natural-language success criteria.\n
- Boundary markers: No specific boundary markers or delimiters are provided in the instructions to isolate user-supplied task strings within the model prompt.\n
- Capability inventory: The skill is capable of executing subprocesses via uv run, cloning repositories via git, and performing file system writes to save annotated HDF5 files.\n
- Sanitization: Input data from recording frames and task strings is not sanitized or validated for embedded instructions before being passed to the VLM endpoint.
Audit Metadata