AAIA · Question #85
A generative AI system has a validation control in place to reject inappropriate questions by checking them against built-in ethical standards. Which of the following enables malicious actors to circu
The correct answer is B. Presenting theoretical situations to justify the reason for asking the questions. Framing an inappropriate request within a hypothetical or theoretical scenario is a well-known jailbreak technique (e.g., 'In a fictional story...', 'Hypothetically speaking...'). The model's ethical filter evaluates the surface framing rather than the underlying harmful intent,
Question
A generative AI system has a validation control in place to reject inappropriate questions by checking them against built-in ethical standards. Which of the following enables malicious actors to circumvent this control through prompt engineering?
Options
- ASubmitting the same questions in a foreign language translated by another AI-based system
- BPresenting theoretical situations to justify the reason for asking the questions
- CAsking the same questions later when the algorithm has changed after further learning
- DRandomly placing keywords unrelated to the main topic
How the community answered
(59 responses)- A7% (4)
- B85% (50)
- C3% (2)
- D5% (3)
Explanation
Framing an inappropriate request within a hypothetical or theoretical scenario is a well-known jailbreak technique (e.g., 'In a fictional story...', 'Hypothetically speaking...'). The model's ethical filter evaluates the surface framing rather than the underlying harmful intent, so the request appears acceptable and passes the validation check. This is a core category of prompt injection attack. Submitting questions in a foreign language (A) may bypass some filters but is less reliable and easily countered. Waiting for the algorithm to change (C) is speculative and not a deliberate exploit. Inserting random unrelated keywords (D) is an obfuscation attempt but does not reliably reframe the intent to pass ethical validation.
Topics
Community Discussion
No community discussion yet for this question.