4.1.2. Real-Time Protection for Microsoft Copilot Studio Agents
💡 First Principle: A Copilot Studio agent built by a business user (not a security team) can still reach sensitive systems and data through its configured connectors — real-time protection exists to apply security oversight to agent behavior as it runs, independent of how carefully (or not) the agent was originally built.
Enabling and configuring real-time protection for Copilot Studio agents means security policies are evaluated against agent actions and outputs as they happen — not just at build time — catching risky behavior (an agent attempting to access data outside its intended scope, or exhibiting behavior consistent with a prompt injection attempt) that a one-time review during agent creation would miss entirely.
⚠️ Exam Trap: Reviewing an agent's configuration once at build time is not equivalent to real-time protection — an agent's behavior can be manipulated at runtime through the inputs it receives (prompt injection), which a static, point-in-time review has no way to catch.
Reflection Question: Why does prompt injection specifically require runtime protection rather than build-time review, unlike a traditional application vulnerability that a code review might catch before deployment?
