Preventing AI Agents from Reading Clinician Teams Chats Through the Microsoft 365 MCP Server
Control agent access to Teams chats in three layers. First, narrow the Entra and Graph permissions the agent's identity holds. Second, expose only the tools the task needs on the MCP server. Third, enforce policy at runtime on each tool call and response, with allow, redact, block or escalate decisions. Retain per-call evidence to support audit controls under HIPAA 164.312(b).
The layers that can restrict access
- 1
Layer 1: Entra and Graph permissions
Grant the agent identity the narrowest chat-related permission that supports the task. Prefer delegated access where the user's own reach is the right ceiling. Treat any application-level grant as a tenant-wide decision requiring explicit approval.
- 2
Layer 2: MCP server tool configuration
Expose only the tools the use case needs. If the task is scheduling, a chat-reading tool should not be visible to the agent at all.
- 3
Layer 3: Runtime policy on tool calls
Evaluate each call and each response in the execution path. Decide to allow, redact, block or escalate based on identity, tool, parameters and content.
Evidence to retain for each tool call
- The user, agent and organization identity behind the call
- The tool invoked and its parameters
- The policy decision: allow, redact, block or escalate
- How the response was handled, including which fields were redacted
- Any human approval, with approver and timestamp
- Retention period and an export path for audit review
How an agent reaches Teams chat content
Model Context Protocol (MCP) is a standard for how applications provide context to large language models. An MCP server publishes a set of tools. The agent decides which tool to call and with what parameters, and the server executes the call against Microsoft Graph using a token. If one of those tools can read chats, the agent can read chats, subject only to the permission behind the token.
That permission is either delegated, acting with the signed-in user's access, or application, acting with a tenant-level grant. In both cases the agent's practical reach is the combination of the tools exposed and the permissions granted. The model chooses which calls to make within that reach, and in a healthcare tenant clinicians routinely discuss patients in chats. Teams meeting chats also persist, so anyone invited to a meeting keeps access to its chat after leaving the meeting. That widens the set of conversations a single grant can touch.
Two documented agentic threat patterns apply directly. Agents can misuse legitimate tools through prompt manipulation or privilege control, which can lead to data exfiltration. Agents can also inherit or delegate credentials across task boundaries, so tokens may propagate and allow access beyond the original scope. A chat-read tool that behaves exactly as designed is therefore still a risk.
Where native controls run out
Native Microsoft 365 permission scoping is coarse relative to conversation content. A permission grant says an identity may read chats. It does not say which chats, for what purpose, or whether the returned text contains patient identifiers. Once the grant exists, any call the tool layer permits will succeed, whether the agent made it for the intended task or because injected instructions pushed it there.
Platform controls such as DLP policies and Microsoft configuration are cited in federal agency assessments as the means of limiting how Copilot handles PII and PHI. They are a sound baseline. They generally operate on configuration and content classification, not on the intent of an individual tool call made by an autonomous agent. Those same assessments describe access to Teams communications in terms of authorized users and administrators, and agents are not addressed. The safe assumption is that an agent is a separate access principal that needs its own limits.
This is why the third layer matters. The tool call is where the decision to read a chat is actually made, so it is the point where scope, purpose and content can be checked together.
Implementation sequence
Work from the outside in, so that each step narrows what the next one has to handle.
- 1. Inventory agent identities and grants: List every agent or MCP server identity in the tenant and the chat-related Graph permissions each holds, delegated or application. Remove anything the documented use case does not require.
- 2. Define data scope per use case: State which resources and time windows each agent may reach. Written scope, applied by analogy to chats, follows the practice of limiting the resource types and time windows AI components can access.
- 3. Trim the tool surface: Disable chat-reading tools for agents that do not need them. Separate agents with different purposes so one token is not reused across tasks.
- 4. Write call-level policy: Express rules on tool, parameters and identity. Examples: block reads of chats outside an approved set, redact patient identifiers in responses, escalate bulk reads to a human.
- 5. Add blast-radius limits: Apply rate limits and throttles so a manipulated agent cannot enumerate many chats quickly.
- 6. Test with adversarial content: Place instructions inside chat messages and confirm that policy, not model behavior, stops the unintended read.
Enforcing policy at the MCP layer with Trussed
Trussed is one way to implement the third layer. It enforces policy in the execution path across models, agents, tools and MCP servers, at the level of individual interactions and tool calls. For this use case, that means a chat-read call can be allowed, redacted, blocked, rate limited or escalated for human approval at runtime, not reviewed after the fact. It does not replace Entra, Graph or MCP server configuration. It covers the gap those controls leave once a permission exists.
Enforcement produces continuous audit-ready evidence, mapped to HIPAA among other frameworks, so the record of what was allowed and blocked comes from the enforcement itself. Trussed can be deployed self-managed in the customer's cloud, with configurable retention and data residency. It runs in production with healthcare and pharma organizations, and has 18+ months of continuous production in a regulated enterprise. Trussed does not claim business transaction identity or end-to-end tracing of agentic workflows.
Three layers between an agent and a clinician chat
Identity and permissions
Entra and Graph grants define the outer limit of what a token can read.
MCP tool configuration
Only the tools a task needs are exposed to the agent.
Runtime policy
Each tool call and response is evaluated: allow, redact, block or escalate.
Evidence
Every decision is recorded and attributable to a user and an agent.
Review your agent access to clinician chats
Walk through your Microsoft 365 MCP configuration and see how runtime enforcement and audit-ready evidence apply to Teams chat access.
Request a DemoHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.