Question 530
A customer service agent may use a tool to cancel real orders but evaluation never tests unauthorized tool calls. A release test must show that an unauthorized cancellation attempt is rejected before the action. What is the best next action for this Azure AI workload?
Explanation
Correct Answer
Tool-safety evaluation must cover whether the agent respects authorized boundaries. A successful run shows that an unauthorized cancellation attempt is rejected before the action.
Incorrect Answers
The alternative is relevant when an agent obediently repeats instructions embedded inside retrieved documents to reveal its secret configuration; tool-safety evaluation must cover whether the agent respects authorized boundaries.
That solution is intended for cases where a production chat agent accepts a hostile user message that asks it to ignore earlier data use restrictions. Here, tool-safety evaluation must cover whether the agent respects authorized boundaries.
Safety controls can require inspection at separate model intervention points.