To prevent agents from obeying malicious instructions hidden in external data, all text entering an agent's context must be ...