i4h-workflow-train-rl
Pass
Audited by Gen Agent Trust Hub on Sep 9, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: Clones the official Isaac for Healthcare workflows repository from GitHub (github.com/isaac-for-healthcare/i4h-workflows) to establish the necessary reinforcement learning environment.
- [COMMAND_EXECUTION]: Executes shell scripts including train.sh and run.sh from the local repository to perform reinforcement learning tasks, model evaluation, and artifact export.
- [INDIRECT_PROMPT_INJECTION]: The skill processes declarative configuration files to define scenes and training objectives. * Ingestion points: Reads YAML profiles and trainer configurations from the ./rl/profiles/ and ./rl/config/ directories (SKILL.md). * Boundary markers: Employs structured YAML for configuration data; explicit instructions to ignore embedded prompts are not included. * Capability inventory: Includes shell command execution, filesystem writes for checkpoint management, and network operations for repository cloning (SKILL.md). * Sanitization: The implementation relies on task-specific adapters and registry-level validation for data consistency.
Audit Metadata