output-validation-sandbox
Installation
SKILL.md
Model Output Validation Sandbox
Security Recommendation
A model response evaluation sandbox MAY be deployed to validate model outputs by performing checks that carry higher risk and must be done in isolation. This is especially relevant when model outputs are executed or trigger actions.
Use Cases
| Scenario | What the Sandbox Does |
|---|---|
| Generated source code | Executes the code in isolation to validate it is safe and correct |
| Generated API calls | Validates the target endpoint and parameters are safe before executing |
| Generated commands | Tests OS commands in a contained environment before execution |
| Generated configurations | Validates configurations won't introduce security misconfigurations |
Critical Requirement: True Isolation
The sandbox MUST be truly isolated so that it does not represent a risk for the system. A compromised sandbox must not be able to: