Responsible AI for generative and agentic systems
AI-103 · Lesson 4 of 15 · 23:52
YouTube loads only after you press play. This chapter opens at its timestamp in the full course. Watch on YouTube ↗
In this lesson
Prepare for the AI-103 exam (Azure AI Apps and Agents Developer Associate). In this lesson: Foundry guardrails (risks, severity thresholds, intervention points for agents, the default policy, content_filter errors), risk detection (Prompt Shields for user prompt and document attacks, protected material, PII, groundedness, task adherence, blocklists), evaluators and the AI Red Teaming Agent (quality and safety scores, defect rate, Attack Success Rate), and audit and governance for agents (tracing, logs, Purview, Content Credentials, MCP approvals, least privilege). Includes three original practice questions.
Preview the topics
- An agent reads web pages with a tool. One page hides a line telling the agent to email customer data to an outside address. Which protection targets this attack?
- A team adds the hate and unfairness evaluator to continuous evaluation, so that hateful answers never reach users. Does that achieve their goal?
- An agent connects to a remote MCP server whose tools can delete customer records. The team wants a person to confirm every call before it runs. What should they configure?