Developers often treat agent harnesses as neutral wiring, but new red-teaming research shows that your choice of harness can make or break your AI security.