Pular para o conteúdo principal

Configure Inline Prompt Injection Protection with Global Secure Access

Implementation Effort: Medium - Configure acquisition, TLS inspection, prompt rules, conversation schemes, and policy assignment.
User Impact: Medium - Detected malicious prompts and rate-limited requests can be blocked; test false positives and business workflows.
Documentation checked: 2026-11-11

Overview​

Prompt Injection Protection in Microsoft Entra Internet Access inspects supported outbound prompt requests and can block detected adversarial prompts and jailbreak attempts before they reach the destination application.

This adds a network defense independent of application code. It does not guarantee detection of every prompt injection, protect every document/retrieval path, or replace application-side safety and least-privilege tool authorization.

This supports Assume breach by adding an independent inspection layer and Verify explicitly by evaluating the supported prompt content rather than trusting an AI application's own controls.

Prerequisites​

  • Complete AI_163 readiness; verify Internet Access acquisition for the intended applications.
  • Internet Access licensing, Global Secure Access Administrator, and Conditional Access Administrator access.
  • The supported setup described in the linked feature guide uses Windows devices/VMs that are Entra joined or hybrid joined. Confirm support before extending beyond that setup.
  • A trusted TLS inspection certificate and policy covering the target HTTPS traffic.

Implementation guidance​

  1. Verify actual forwarding before configuring prompt protection.
  2. Under Global Secure Access > Secure > Prompt policies, create a policy and a rule with the intended status/priority and Block action.
  3. Add a conversation scheme matching the target application's request format. Use available preconfigured extractors or AI_031 for custom JSON applications.
  4. Link the prompt policy to a new or existing GSA security profile.
  5. Create a Conditional Access policy for the pilot users/groups targeting All internet resources with Global Secure Access. Under Session, select Use Global Secure Access Security Profile and the intended profile.
  6. Review configuration and enable enforcement for the pilot; report-only assignment is not evidence of blocking.
  7. Run controlled benign and adversarial test prompts approved for the test environment. Expand only after verifying decisions and application functionality.

Validation and completion evidence​

  • A normal prompt succeeds and an approved prompt-injection test is blocked.
  • Evidence identifies target endpoint, policy/profile assignment, and the matching network decision.
  • Owners can investigate false positives and rate-limit blocks and maintain a rollback/exception process.

Scope and limitations​

  • Supports text prompts in JSON-based generative AI apps; files are not supported.
  • Prompts beyond 64,000 characters are truncated.
  • Scanning is rate limited; requests exceeding the limit are blocked.
  • Prompt extraction and application compatibility must be validated; a brand name does not prove every API or UI route is covered.
  • Copilot Studio's GSA integration excludes its own LLM orchestration/result-enhancement requests. Enabling that integration is not evidence that those calls receive this protection.

Reference​