We introduce SandboxEscapeBench, the first benchmark to systematically evaluate whether AI agents can break out of their sandboxes, and share some early results.

A comprehensive toolkit for safely evaluating AI agents.