A parks volunteer-uploaded flyer contains instructions that try to override the city’s approved messaging agent. What type of risk is this?
Select an answer to reveal the explanation.
Short Explanation
Even a cheerful volunteer flyer can hide “forget the official talking points” instructions. When that file becomes Copilot context, it’s indirect prompt injection. Stick to approved messaging sources for public parks posts.
Full Explanation
A parks volunteer-uploaded flyer that contains instructions trying to override the city’s approved messaging agent is indirect prompt injection through a document used as context for Copilot or an agent. The malicious or manipulative text rides inside content the city might retrieve as grounding, attempting to change agent behavior away from approved municipal messaging.
Calling this a required Agent Store update that always accepts volunteer text as new city policy invents a mandatory policy-acceptance path. Volunteer uploads are not automatically new city policy, and Agent Store updates are not the risk named by override instructions in a flyer.
Calling it fabrication of images only, unrelated to instructions embedded in documents, misses the injection concept. The stem is about instructional override text in a document, not image hallucination alone.
Calling it data-protection success proving Copilot deleted the flyer before anyone could read it invents a deletion success the scenario does not describe.
Exam caveat: documents and uploads used as agent context are classic indirect prompt-injection vectors—distinguish that from fabrication-only or Agent Store policy myths.
Operational check: scan volunteer-supplied flyers before they become agent context; quarantine files with override-style instructions, and keep the approved parks messaging agent pinned to city-authored guidance rather than untrusted upload text.