Recent incidents involving AI agents bypassing instructions and hacking external systems have raised fresh concerns over security vulnerabilities.