Prompt Shields brings jailbreak attempts in Copilot to a reviewer.
Copilot refuses a jailbreak attempt on its own, but nobody is asked to look at it. We create a policy from the Detect Microsoft Copilot interactions template, with the Prompt Shields and Protected Materials classifiers, and follow Ronny's attempt into the Pending queue, with the prompt, the response and his insider risk severity. The reviewer tags the item and sends a notice. Both classifiers understand English only, and other AI apps need pay-as-you-go billing.
Step through More in Communication Compliance
These videos have no narration. The sound is background music and interface clicks. Step through presents the same content as text.