Plan and manage an Azure AI solution 1.4 Implement responsible AI across generative AI and agentic systems
Before release, a safety team evaluates a procurement agent against a user-verified…
Study guide skill: Apply responsible AI instrumentation, including evaluators, safety evaluations, and explanation tooling
Question B8-011
Before release, a safety team evaluates a procurement agent against a user-verified policy that lists tool uses the agent must never perform, such as approving payments above its authority. Which built-in risk and safety evaluator fits?
Select one answer.
Answer: A
Prohibited actions (preview), which supports agents only, measures whether an agent engages in explicitly disallowed actions or tool uses based on a user-verified policy. Sensitive data leakage measures exposure of sensitive information. Ungrounded attributes looks for ungrounded inferences about personal attributes, and Code vulnerability looks for insecure generated code.
Deep dive
How it works
Foundry's risk and safety evaluators each target one kind of risk, and some apply only to agents. Prohibited actions (preview) is agents-only. It measures an AI agent's ability to engage in behaviors that violate explicitly disallowed actions or tool uses, based on a user-verified policy of prohibited actions.
The safety team has exactly that: a verified list of tool uses the procurement agent must never perform, such as approving payments above its authority. Prohibited actions checks the agent's behavior against that list.
How the other evaluators in the question differ:
- Sensitive data leakage (preview), also agents-only: an agent's vulnerability to exposing sensitive information such as financial data, personal identifiers or health data.
- Ungrounded attributes, for models and agents: text that contains ungrounded inferences about personal attributes, such as demographics or emotional state.
- Code vulnerability, for models and agents: generated code with security vulnerabilities, such as code injection or SQL injection.
The other options, and when you'd use them
- Ungrounded attributes: For ungrounded inferences about people's personal attributes, not tool-use policy.
- Code vulnerability: For insecure generated code. The procurement agent's risk is a forbidden action, not bad code.
- Sensitive data leakage: For exposing sensitive information. Approving an oversized payment is a disallowed action, not a data leak.
Exam tip
Policy list of forbidden tool uses = Prohibited actions (agents only). Exposing secrets or PII = Sensitive data leakage.
References
Last reviewed: