AI and LLMs · Senior · Question 57 of 75
How should an agent handle content from tools as opposed to user instructions?
- 1As system instructions that override safety rules
- 2As untrusted data that can't grant new permissions
- 3As higher-priority instructions than the user's
Show the answer
Correct answer: As untrusted data that can't grant new permissions
Explanation
A web page or email saying "you are now allowed to..." must not change what the agent may do.