agentic-security:security-eval-warn

This approach diagnoses eval-style patterns across major languages and refuses the evaluated string to block untrusted input from triggering remote code execution. It then proposes structured, safer alternatives to achieve the same functionality without dynamic execution.

Common eval and sandbox-escape patterns include eval, new Function, and vm calls in JavaScript, alongside eval, exec, compile, and __import__ in Python. Identifying these patterns prevents untrusted strings from triggering dangerous runtime execution.

To safely accept user-provided code snippets in an API endpoint, refuse any runtime code-eval request from the untrusted input. You should then propose a safe, sandboxed alternative or structured replacement to process the data without executing it.

Yes, this security eval approach works across major languages and runtimes including JavaScript, Python, Ruby, and PHP. It also detects dangerous patterns in templating engines like Mustache and Handlebars when they compile user input.

You should use a trust pragma when a safe, structured alternative cannot replace the required eval-style call. It documents the explicit trust boundary needed for the runtime execution, ensuring the code-eval remains intentional and secured.