execute-script-safely

To safely execute ephemeral scripts, you should enforce a security pre-review using a haiku Agent to evaluate the code before running it. This process gates script execution by providing a clean, suspicious, or dangerous verdict to prevent unsafe actions and potential data exposure.

The best way to validate sub-agent generated scripts is to run them through a haiku Agent pre-review workflow. This evaluates the temporary snippet and returns a verdict, allowing you to execute clean scripts, stop dangerous ones, or prompt the user for suspicious content.

You need a sandbox to run temporary investigation snippets because it blocks external exfiltration while allowing safe inspection of code behavior. However, you must also apply a pre-review step because the sandbox cannot guarantee safety against reading local secrets.

The limitations of using a sandbox for script execution security include an inability to guarantee safety against reading local secrets or CWD-level code poisoning. It successfully blocks external exfiltration but requires a pre-review workflow to mitigate local data exposure risks.