The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructions and scripts facilitate downloading code and configuration from an external source.
Evidence: git clone https://github.com/ShandaAI/WildWorld.git in SKILL.md.
[REMOTE_CODE_EXECUTION]: The skill recommends installing dependencies from a remote requirements.txt file located in the cloned repository, which could execute arbitrary code during installation.
Evidence: pip install -r requirements.txt in SKILL.md.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes untrusted data from external dataset files (JSON annotations and metadata) without explicit sanitization or boundary markers.
Ingestion points: Data is read via json.load() in get_skeleton, get_state, get_camera, and get_action methods within SKILL.md.
Boundary markers: Absent. The skill does not implement delimiters or instructions to ignore embedded commands in the dataset files.
Capability inventory: The skill performs file system reads, image processing with PIL, and tensor operations with PyTorch.
Sanitization: Absent. The skill uses standard json.load() which does not filter for potential prompt injection content in the natural language fields (captions/actions).

wildworld-dataset