Extend Prompt Injection Protection to Custom JSON-Based AI Endpoints
Implementation Effort: Medium - Map each supported application's request format and validate extraction and enforcement.
User Impact: Low - Existing approved requests should remain functional; malformed mappings or detections can cause blocks.
Documentation checked: 2026-11-11
Overview
Custom enterprise AI applications often wrap model APIs or expose prompts through a different JSON schema. A built-in conversation scheme may not extract the prompt correctly from those requests.
Global Secure Access (GSA) supports a Custom conversation scheme using the destination service URL and the JSON path of the prompt in the request body. This extends the prompt policy established in AI_029 to supported custom JSON-based applications.
This supports Verify explicitly by validating the actual request field being inspected and Assume breach by reducing blind spots at custom application boundaries.
Prerequisites
- Complete AI_029, including the documented client, forwarding, TLS, profile, and Conditional Access prerequisites.
- An application owner who can provide the endpoint URL, request schema, and safe test environment.
- Confirm that traffic to the endpoint is acquired by Internet Access and is eligible for this feature.
Implementation guidance
- Inventory the custom applications and endpoints that send prompts. Obtain representative requests using synthetic data.
- Determine the exact service URL and JSON path containing the prompt. Map every relevant supported request route; do not assume one route represents the entire application.
- In the prompt rule's conversation scheme configuration, select Custom and enter the URL and prompt JSON path.
- Add the scheme to the intended prompt policy. Verify that the policy remains linked to the active security profile and scoped to the correct pilot users.
- Compare benign requests and approved adversarial test requests to prove extraction and blocking.
- Make request-schema and endpoint changes trigger retesting. Record application owner, scheme owner, last test date, and exceptions.
Validation and completion evidence
- Every claimed protected endpoint has a documented request mapping and a correlated test.
- Benign requests succeed; an approved adversarial prompt in the mapped field is blocked.
- Alternate routes or request formats are either tested or recorded as outside coverage.
- Schema-change monitoring and rollback are assigned.
Scope and limitations
The documented custom configuration maps the request prompt, not an arbitrary model-response JSON schema. Do not claim general response inspection from this control.
The same limits as AI_029 apply: JSON-based applications, text prompts only, truncation beyond 64,000 characters, and rate-limit blocking. Private endpoints are not automatically covered by Internet Access or by adding a scheme. Non-JSON, multimodal, file, and unacquired traffic require other controls.