The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides instructions that download a remote repository via git clone and then execute a script (viz.sh) from within that repository. Because the source URL is derived from the I4H_WORKFLOWS_REPO_URL environment variable, an agent could be directed to execute arbitrary code if this variable is pointed to an untrusted location.
[EXTERNAL_DOWNLOADS]: The skill performs an external download of the 'Isaac for Healthcare' workflow tools from GitHub. While this is a vendor-associated resource, it involves fetching external code at runtime.
[COMMAND_EXECUTION]: The skill executes local shell scripts and commands (e.g., viz.sh, find, git) to search for datasets and launch a web server.
[INDIRECT_PROMPT_INJECTION]: The skill processes and reports on metadata (info.json) and task descriptions stored within local datasets. A malicious dataset could contain instructions designed to influence the agent's summary or subsequent actions.
Ingestion points: Reads info.json and metadata from directories in runs/ or ~/.cache/huggingface/lerobot/.
Boundary markers: None identified.
Capability inventory: File system searching, metadata reading, and shell script execution.
Sanitization: No evidence of sanitization or validation of the ingested metadata before processing.
[DYNAMIC_EXECUTION]: The skill dynamically determines its execution paths and the source of its tools using environment variables and string manipulation, which increases the complexity of verifying the code that will ultimately be executed.

i4h-lerobot-viz