AI agents can escape sandboxes without ever breaking them

Sandboxes have become a key security control for AI coding agents, but new research suggests they may not provide the isolation many organizations assume. Pi...