skill-sandbox
Pass
Audited by Gen Agent Trust Hub on Sep 7, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill workflow involves executing local validation scripts, including validate-skills.mjs, audit-semantics.mjs, and evaluate-skill, to test experimental code within the project workspace. These are standard development operations for skill auditing and validation.
- [INDIRECT_PROMPT_INJECTION]: The skill processes user-provided skill concepts and capabilities to generate new instruction files (SKILL.md). While this provides a surface for instruction injection, the risk is mitigated by the isolated sandbox structure and the requirement for formal validation steps prior to any promotion to the core skill bundle. • Ingestion points: Processes skill concepts and desired capabilities from user input to build the SKILL.md body. • Boundary markers: Explicitly isolates all experimental work in a .skill-sandbox/ directory to prevent impact on canonical skills. • Capability inventory: Performs file system writes (creation of sandbox), script execution (validation), and conditional file moves (promotion). • Sanitization: Relies on the validation scripts (validate-skills.mjs) to enforce platform schemas and semantic standards on the generated content.
Audit Metadata