Cisco will use Claude Mythos to help businesses test AI agents for security weaknesses. Its safeguards sound sensible, but effectiveness remains to be demonstrated.