Microsoft Copilot Studio
Alice connects to Microsoft Copilot Studio as an Environment Security Provider (ESP), Microsoft's external threat-detection interface. Enabled once per Power Platform environment, it covers every agent in it: Copilot Studio calls Alice as an agent is about to run a tool, and a tool Alice blocks is not run. Alice evaluates the user message that led to the tool and, with tool message types enabled on your project, the tool call itself (the tool and the arguments it is about to receive) and the outputs of the tools that already ran in the plan. Agent responses are not evaluated: Copilot Studio does not send them to security providers.
Status: Generally available · Evaluates: Prompts, Tool calls · Vendor: Microsoft
The user message on every tool run; the tool call and earlier tool outputs with tool message types. Copilot Studio does not send agent responses.
Setup
Configuration: Entra tenant id (saved in Alice), Alice Copilot protection app registration
- An Entra admin grants organisation consent to the Alice Copilot protection application via the consent URL shown in Alice under Account Settings, Copilot Integration.
- In Alice, add your Entra tenant id(s) on that page and save.
- An Entra application admin registers the app with Microsoft's
Create-CopilotWebhookApp.ps1script (PowerShell Gallery), pointing it at the Alice endpoint, or manually with the federated credential values shown on the page. - In the Power Platform admin center, configure the threat detection system for the environment. Every agent in it is covered from then on.
Good to know
Enabled per project by Alice on request. Applies to generative-orchestration agents. Verdicts are allow or block; there is no mask. Each agent gets its own application in Application Inventory the first time it sends traffic.
Tool calls and tool outputs are evaluated only once Alice has enabled WONDER_FENCE_AGENTIC_MESSAGE_TYPES on your project, and only by policies that list the Tool Call or Tool Result message type. A policy that lists only Prompt does not screen them. Each output is screened until it passes once in the conversation, and up to the eight most recent outputs are read per tool run. Copilot Studio runs the tool if Alice has not answered within 1,000 ms.