انتقل إلى المحتوى الرئيسي

Verify Global Secure Access Readiness for AI Workloads

Implementation Effort: Low - Validate existing infrastructure and identify prerequisites; deployments discovered to be missing are separate work.
User Impact: Low - Assessment only; routing or enforcement changes require a controlled rollout.
Documentation checked: 2026-11-11

Overview​

AI network protection depends on a verified traffic path and the correct enforcement policy for each traffic source. Microsoft Entra Internet Access can protect supported endpoint, remote-network, and Copilot Studio traffic, but those sources do not share identical prerequisites or feature support.

Traffic acquisition forwards connections to Global Secure Access (GSA); it does not itself enforce an Allow or Block decision. Likewise, policy creation is not evidence that a workload is protected. Administrators must establish acquisition, policy attachment, supported inspection, and runtime evidence.

TLS inspection is required for encrypted payload inspection such as prompt protection, client-path MCP inspection, and network content inspection. It is not a universal prerequisite for every destination control: SNI-based HTTPS destination filtering can apply without decryption. Copilot Studio MCP logging also has a documented integration path without separately configured TLS inspection.

This supports Verify explicitly by proving coverage before enabling AI workloads and Assume breach by identifying bypasses and unsupported traffic paths rather than treating configuration as protection.

Implementation guidance​

  1. Inventory AI traffic sources: user applications, local agents, Copilot Studio environments/connectors, branch traffic, and private workloads. Identify destinations, ports, protocols, identity context, and owners.
  2. Confirm licensing, administrative roles, supported client versions, user/group assignments, environment prerequisites, and tenant availability for each chosen feature.
  3. Review Internet Access forwarding, Default/Custom/Agentic Acquire choices, bypass rules, other SWG/VPN coexistence, and Private Access routing. Select acquisition by source; do not enable every mode indiscriminately.
  4. For TLS-dependent features, verify certificate trust, inspection policy coverage, and approved exclusions. Account for certificate-pinned applications and unsupported transport paths.
  5. Confirm shared destination, threat-intelligence, and basic content controls with the network team. Agree AI-specific rules and exceptions.
  6. With the Purview team, determine whether Scan with Purview is applicable, licensed, and mapped to an Entra user identity. Do not assume autonomous agent-only traffic is scanned.
  7. Record unsupported flows and compensating controls, including private endpoints and Copilot Studio exclusions. Establish monitoring, change ownership, and rollback.

Validation and completion evidence​

  • A source-to-destination coverage matrix identifies acquisition and enforcement separately.
  • A pilot request is visible in GSA diagnostics/logs with the expected source and destination.
  • Each planned protection has a supported source, inspection prerequisite, policy attachment, and test owner.
  • Exceptions and residual risks are documented; missing prerequisites have tracked remediation.

Scope and limitations​

This is a readiness checkpoint, not a declaration that all AI traffic is protected. Use Agentic Acquire for applicable local agents and Copilot Studio forwarding for supported hosted-agent interactions. Private Access and Azure private networking require separate assessment.

Reference​