Researchers escaped four top AI coding agents’ sandboxes without ever breaking them