CrowdStrike says AI agents can take unauthorized actions even without a hacker involved. Its new kill switch gives security teams a way to intervene, with endpoint termination still to come.