Breaking out: Can AI agents escape their sandboxes?

SandboxEscapeBench is an open-source benchmark from the University of Oxford and the AI Security Institute that tests whether AI agents in container sandboxes can escape to a host filesystem to retrieve a protected file. The benchmark covers 18 scenarios across orchestration, runtime, and kernel layers and focuses on known vulnerability classes (e.g., exposed Docker sockets, writable host mounts, privileged containers, Dirty COW/Dirty Pipe). Results show frontier models can exploit common misconfigurations but failed on more complex kernel-level exploits; all successful escapes used known, publicly disclosed weaknesses and no new vulnerabilities were discovered.