Which prompting approach can help defend against prompt-injection attacks?

Adversarial prompting is a technique where you intentionally craft prompts that attempt to bypass or manipulate the model's safety mechanisms, similar to how an attacker might. By doing this in a controlled environment, you can identify vulnerabilities and then develop more robust defenses against prompt injection attacks. Zero-shot, least-to-most, and chain-of-thought prompting are all techniques to improve a model's performance or reasoning abilities for legitimate tasks, not primarily to defend against prompt injection. While they might indirectly make a model more robust by improving its understanding, they are not designed as a direct defense mechanism against malicious prompts.