Skip to content

Latest commit

 

History

History
34 lines (25 loc) · 1.85 KB

File metadata and controls

34 lines (25 loc) · 1.85 KB

Threat Model

Assets

  • Prompt content and workflow results
  • CAS lifecycle metadata and traces
  • Foundry project and agent
  • Container App system-assigned identity

Trust Boundaries

  • Caller to public API
  • Application to Foundry project endpoint
  • Application telemetry to Azure Monitor
  • CI to public package and container registries

Threats and Controls

Threat Control Residual risk
Credential disclosure No keys or tokens in code; system-assigned ManagedIdentityCredential in Azure Operators must maintain least-privilege RBAC
Legacy API use Adapter uses project Responses client with agent_reference; no Classic Assistants code SDK behavior must be reviewed during upgrades
Prompt injection Workflow treats prompts as untrusted data and exposes no tools in v0.1 Downstream agent policy remains product-specific
Sensitive telemetry Explicit spans contain safe identifiers only; broad outbound HTTP/SDK auto-instrumentation and retry-file storage are disabled Operators must review platform log settings
Unauthorized invocation External ingress disabled by default in platform interface Authentication gateway is product-specific and out of scope
Supply-chain compromise Pinned CI actions, lint, tests, non-root container Dependency update review remains required
Denial of service Platform scaling bounds and request validation Product-specific quotas and rate limits are not included

Required RBAC

Grant the Container App system-assigned identity only the minimum Foundry project role needed to invoke the selected agent, scoped to the narrowest resource. Do not assign subscription-wide roles.

When Application Insights export is enabled, also grant the minimum Azure Monitor publishing role required by the deployment at the narrowest telemetry resource scope.