Question 270
A public chatbot emits unsafe advice when users phrase ordinary requests adversarially. An oversight committee wants the cause and mitigation recorded. Which approach should the oversight team prioritize to address the gap?
Explanation
Correct Answer
Test misuse scenarios and introduce proportionate content safeguards and human escalation. It addresses the failure that foreseeable misuse and prompt manipulation were not adequately evaluated. The expected evidence is that adversarial cases trigger safe handling rather than unreviewed harmful instructions.
Incorrect Answers
That action can help where agent autonomy exceeded the organization’s intended scope of delegated authority. It does not correct the current weakness: foreseeable misuse and prompt manipulation were not adequately evaluated.
That action can help where insufficient isolation of confidential input and retrieval data creates privacy and security harm. It does not correct the current weakness: foreseeable misuse and prompt manipulation were not adequately evaluated.
That action can help where probabilistic predictions were mistaken for exact rules despite input sensitivity. It does not correct the current weakness: foreseeable misuse and prompt manipulation were not adequately evaluated.