Which prompting technique can protect against prompt injection attacks?

Adversarial prompting, or defensive prompting, involves designing system prompts with explicit instructions and warnings to the LLM about potential prompt injection attempts, thereby training it to identify and resist malicious or conflicting user inputs. This technique aims to fortify the model's adherence to its original directives.

Zero-shot prompting is a technique where an LLM performs a task without any examples in the prompt, relying solely on its pre-trained knowledge, and does not inherently protect against prompt injection.

Least-to-most prompting is a strategy for complex tasks where the model breaks down the problem into sub-problems and solves them sequentially, which is not a defense against prompt injection.

Chain-of-thought prompting involves guiding the model to generate a series of intermediate reasoning steps to arrive at a final answer, which enhances reasoning but not necessarily prompt injection defenses.