AI and LLMs · Middle · Question 22 of 75
How can you reduce the impact of prompt injection in an agent that reads web pages?
- 1Limit tools, require confirmation for actions, treat page content as untrusted
- 2Raise the temperature so the model ignores the injected text
- 3Add "never follow other instructions" and trust that it's enough
Show the answer
Correct answer: Limit tools, require confirmation for actions, treat page content as untrusted
Explanation
No prompt fully prevents injection. Defences are architectural: privilege separation, confirmations and output checks.