AI-300 Practice Question 59: AI Red Teaming Agent

Microsoft Machine Learning Operations Engineer Associate (AI-300) · Test 6 · question 9 of 10

Implement generative AI quality assurance and observabilityhard

Before launching a customer-facing agent, a security team wants an automated and repeatable scan that attacks the target with adversarial prompts by using several attack strategies, and then reports how often the attacks succeed for each risk category. What should they use?

Select one answer
Show answer and explanation

Answer:

  • C. The AI Red Teaming Agent, which reports attack success rate per risk category

The AI Red Teaming Agent uses Microsoft's PyRIT framework to send adversarial prompts to a target, applying attack strategies that transform those prompts. The scan reports an attack success rate by risk category, so the result is a repeatable, measurable pre-release test (option C).

- A is wrong because groundedness measures whether answers are supported by context. It does not generate adversarial attacks. - B is wrong because a content filter is a runtime control that blocks content. It does not probe the system or report how often attacks succeed. - D is wrong because continuous evaluation scores real traffic that already happened. It does not actively send crafted attacks before launch.

Reference: AI Red Teaming Agent

Practice Test 6 ›

Study guides for AI-300