Edit

Data flows and threat model for Microsoft Discovery

This article describes the trust boundaries and data flows in a Discovery deployment. It maps them to threats, built-in mitigations, and the controls that you own. Use it as a starting point for your own threat model. Your network design, workspace configuration, agents, and tools change the flows, so record your actual configuration as evidence.

For the shared responsibility model and the controls you inherit from Azure, see Security and compliance overview and Compliance guidance.

Trust boundaries

Each trust boundary groups components by where they're hosted and who operates them.

Boundary Contents Hosted in Operated by
User device The browser that runs Discovery Studio, and the user's sign-in session Your users' devices You
Microsoft service endpoints Discovery Studio, the Discovery control plane in the home region, and the Discovery MCP Server Microsoft-operated infrastructure, outside your subscription Microsoft
Discovery managed resource groups Backend resources that Discovery creates for each workspace, supercomputer, and bookshelf: Microsoft Foundry, Azure Cosmos DB, Azure AI Search, Azure SQL Database, Azure Storage, Azure Key Vault, Azure Container Apps, Azure Kubernetes Service, and Log Analytics Your subscription Microsoft, through Discovery service principals. Don't modify these resources.
Your resources Resources that you create and connect to Discovery: virtual network and subnets, storage accounts, container registry, Key Vault, managed identities, and firewall Your subscription You
External services GitHub Copilot, Grounding with Bing Search, remote MCP servers, public scientific databases, and package registries Outside your subscription GitHub, Microsoft, or third parties, under their own terms

Discovery managed resource groups and your resources are both in your subscription and Microsoft Entra tenant, so the resources that store customer data stay within your subscription and tenant boundary. The difference is who operates them. Discovery operates the managed resource groups for you, and you operate your own resources. The Azure Storage and Azure Key Vault resources in a managed resource group are separate from the storage accounts and Key Vault that you create. In a cross-region deployment, the home-region control plane handles authorization context, resource metadata, provisioning, and operation status, but it doesn't transport or persist bulk customer data.

Data flows

ID Flow Data Protection
F1 Users sign in to Microsoft Discovery Studio. Credentials and session Discovery Studio is publicly accessible and secured with Microsoft Entra ID.
F2 Discovery Studio and API clients call the workspace and bookshelf data planes. Prompts, results, and metadata Entra ID and Azure RBAC. Optionally, private endpoints with public network access disabled.
F3 The browser reads and writes files in your storage account. Input and output files Direct browser access with CORS rules for Discovery Studio and the Visual Studio Code origins.
F4 The control plane provisions and operates resources in managed resource groups. Configuration and metadata Discovery service principals, and a network security perimeter (NSP) around managed resources.
F5 Agents call chat model deployments. Prompts, retrieved content, and responses NSP and private endpoints, and Microsoft Foundry Guardrails content filtering applied by default.
F6 Agents run tools on the supercomputer, and tools read and write your storage. Input and output files The workload identity and input and output mounts.
F7 Tools call public endpoints and download packages. Queries and identifiers Your firewall and network security groups.
F8 A bookshelf indexes documents and answers queries. Document text, embeddings, and graph The bookshelf workload identity and the bookshelf managed resource group.
F9 Discovery Studio displays results. Summarized results and size-limited previews Entra ID.
F10 GitHub Copilot runs in the workbench. Prompts and file context Requires a GitHub account and the discovery.workbench.enableGhcpAiFeatures tag. Calls reach Discovery through the authenticated Discovery MCP Server. See GitHub Copilot and Microsoft Discovery.
F11 Agents use the web search tool. Search queries derived from the conversation Data sent to Grounding with Bing Search is subject to Bing's terms.
F12 Agents call remote MCP servers. Tool arguments and results Third-party terms and API keys.
F13 Discovery writes logs. Operation traces Application logs in the managed resource group, activity logs in Azure Monitor, and audit logs that you enable.
F14 The supercomputer pulls tool images. Container images The kubelet identity with the AcrPull role.

Threats and mitigations

The following tables apply the STRIDE categories to the data flows. Each row lists the built-in mitigation and the control that you configure.

Spoofing

Threat Flows Built-in mitigation Your control
A stolen user token is used to access Discovery Studio or the data plane. F1, F2 Entra ID authentication Conditional Access, phishing-resistant multifactor authentication, and private endpoints with public network access disabled
A caller other than the Discovery service reaches the supercomputer management plane. F4 IP restrictions to the Discovery service, or a private supercomputer Register the required feature or create a private supercomputer when you create the supercomputer

Tampering

Threat Flows Built-in mitigation Your control
A compromised base image or package enters a tool image. F14 None specific to Discovery Use immutable image digests, a private registry, and image scanning. See Supercomputer AKS upgrades.
Someone modifies resources in a managed resource group. F4 Discovery manages these resources Limit write access to managed resource groups, and exclude them from Azure Policy remediation.
A tool changes data that it shouldn't access. F6 Storage containers scope project data Scope the workload identity's storage role to the storage account.

Repudiation

Threat Flows Built-in mitigation Your control
A change or data access can't be attributed after the activity log retention period. F13 Activity logs retained for 90 days, and correlation IDs Enable audit logging to a Log Analytics workspace or storage account that you own, with a retention policy.

Information disclosure

Threat Flows Built-in mitigation Your control
Code generated by an agent sends data to an arbitrary endpoint. F7 Discovery-delegated subnets have no default outbound route Allow only required FQDNs on a firewall for the node pool subnets.
Conversation content reaches web search or a third-party MCP server. F11, F12 The agent definition decides which tools are available Approve web search and remote MCP agents for each project, and review third-party terms.
Prompts and file context reach GitHub Copilot. F10 GitHub Copilot features require a workspace tag Decide on the tag when you create the workspace. See Use GitHub Copilot in Microsoft Discovery.
Your storage account is reachable from public networks. F3 Microsoft Entra ID authentication Use network-hardened customer storage.
The supercomputer's AKS API server is reachable through a public FQDN. F4 IP restrictions to the Discovery service Create a private supercomputer.
Users with read access to logs see sensitive operation details. F13 Logs stay in your subscription Restrict access to Log Analytics workspaces.

Denial of service

Threat Flows Built-in mitigation Your control
Quota exhaustion blocks workspace or node pool creation. F4 None Plan and monitor quota. See Quota reservations.
AKS upgrades or Spot evictions interrupt long-running jobs. F6 A pod disruption budget protects running tools from voluntary disruption Make jobs restartable, and use Regular priority for jobs that can't restart.
Public APIs limit or block tool requests. F7 None Use API keys where providers offer them, and handle retries in tools.

Elevation of privilege

Threat Flows Built-in mitigation Your control
An agent uses the workload identity to manage Discovery resources. F6 Separate identity slots Assign the workload identity only Storage Blob Data Contributor on the storage account. See Managed identities.
A user with a platform role reaches every project in a workspace. F2 Project-level roles Assign project roles with companion resource roles. See Project-level access control.

AI-specific risks

The following table maps risks from the OWASP Top 10 for LLM Applications to Discovery.

Risk How it appears in Discovery Built-in mitigation Your control
Prompt injection Instructions in web pages, knowledge base documents, or API results steer an agent. Foundry Guardrails content filtering Least-privilege workload identity, restricted outbound access, and human review
Sensitive information disclosure An agent includes private data in a web search, an MCP call, or a shared output. NSP on managed resources Project isolation, storage scoping, and approval of external tools
Supply chain Third-party tool images, packages, and model weights None specific to Discovery Image pinning and scanning, and a private registry
Excessive agency A tool with broad permissions or open outbound access acts beyond the request. Identity slots Separate identities and FQDN-based firewall rules
Misinformation Plausible but incorrect scientific output Grounding with citations Human review of agent and model outputs

For Discovery's responsible AI controls and prohibited uses, see Responsible AI in Microsoft Discovery.