Guardrail is designed for defensive validation, not sandboxing. It does not execute validated values or call __tostring while constructing normal errors. Input values are excluded from serialized errors by default.

For untrusted input, keep max_depth and max_errors bounded, use exact table schemas, and avoid custom validators that execute untrusted callbacks. Validation cannot protect against resource exhaustion caused before Lua receives a value, hostile runtime extensions, or unsafe behavior in an application callback.

Report security concerns privately to the repository maintainers rather than opening a public issue.