The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides instructions for the agent to build and execute complex shell commands using Bash and docker exec to launch experiments locally or within containers. The instructions require the agent to interpolate user-supplied "overrides" directly into the command string. While there is a guideline to quote shell-sensitive characters, this manual escaping process is prone to failure and represents a command injection vulnerability if user input is not handled perfectly by the agent.
[INDIRECT_PROMPT_INJECTION]: The skill requires the agent to read and interpret the content of output files such as arena_experiment_result.json and index.html to verify experiment success and report findings. This creates a surface where malicious content within these artifacts could influence the agent's logic or subsequent commands.
Ingestion points: The agent reads arena_experiment_result.json, index.html, and episode_results_rebuild<N>.jsonl from the output directory to confirm per-run statuses and metrics.
Boundary markers: The skill does not describe the use of delimiters or specific safety instructions to isolate the data in these files from the agent's instruction processing.
Capability inventory: The agent possesses significant capabilities, including shell access to the host and Docker containers, which increases the impact of potential indirect prompt injection.
Sanitization: No sanitization, schema validation, or content filtering is performed on the experiment artifacts before they are analyzed by the agent.
[EXTERNAL_DOWNLOADS]: The skill references and triggers workflows for OpenPI and GR00T policies that involve downloading model checkpoints and building Docker images. While these are part of the intended primary purpose of the skill for the isaac-sim vendor ecosystem, they represent the introduction of external dependencies into the execution environment.

run-experiment