Data Loss Prevention (DLP)

Data Loss Prevention (DLP) for AI Gateway helps protect your organization from inadvertent exposure of sensitive data through AI interactions. By integrating with Cloudflare's proven DLP technology, AI Gateway can scan both incoming prompts and outgoing AI responses for sensitive information, ensuring your AI applications maintain security and compliance standards.

AI Gateway DLP leverages the same powerful detection engines used in Cloudflare's Data Loss Prevention solution to scan AI traffic in real-time. The system analyzes both user prompts sent to AI models and responses received from AI providers, identifying sensitive data patterns and taking appropriate protective actions.

• Prevent data leakage: Stop sensitive information from being inadvertently shared with AI providers or exposed in AI responses
• Maintain compliance: Help meet regulatory requirements like GDPR, HIPAA, and PCI DSS
• Consistent protection: Apply the same DLP policies across all AI providers and models
• Audit visibility: Comprehensive logging and reporting for security and compliance teams
• Zero-code integration: Enable protection without modifying existing AI applications

AI Gateway DLP can scan:

• User prompts - Content submitted to AI models, including text, code, and structured data
• AI responses - Output generated by AI models before being returned to users

The system works with all AI providers supported by AI Gateway, providing consistent protection regardless of which models or services you use.

DLP inspects the text content of request and response bodies as they pass through AI Gateway. The following details apply:

• Non-streaming requests and responses: DLP scans the full request and response body.
• Streaming (SSE) responses: DLP buffers the full streamed response before scanning. This means DLP-scanned streaming responses are not delivered incrementally to the client. Expect increased time-to-first-token latency when DLP response scanning is enabled on streaming requests, because the entire response must be received from the provider before DLP can evaluate it and release it to the client.
• Tool call arguments and results: DLP scans the text content present in the message body, which includes tool call arguments and results if they appear in the JSON request or response payload.
• Base64-encoded images and file attachments: DLP does not decode base64-encoded content or follow external URLs. Only the raw text of the request and response body is inspected.
• Multipart form data: DLP scans the text portions of the request body. Binary data within multipart payloads is not inspected.

When DLP response scanning is enabled and a client sends a streaming request ("stream": true), AI Gateway buffers the complete provider response before running DLP inspection. This differs from requests without DLP, where streamed chunks are forwarded to the client as they arrive.

Because of this buffering:

• Time-to-first-token latency increases proportionally to the full response generation time.
• Request-only DLP scanning (where the Check setting is set to Request) does not buffer the response and has no impact on streaming latency.
• If you need low-latency streaming for certain requests while still using DLP on the same gateway, consider setting the DLP policy Check to Request only, or use separate gateways for latency-sensitive and DLP-scanned traffic.

DLP scanning runs after a cache miss, when AI Gateway forwards the request to the provider and receives a response. The following table describes how each DLP outcome affects caching:

Cache hits skip DLP scanning. When a subsequent identical request matches a cached response, AI Gateway serves it directly from cache without re-running DLP. This is safe because only responses that already passed DLP (or were flagged, not blocked) are cached. However, if you update your DLP policies after a response has been cached, the cached response is not re-evaluated. It continues to be served until the cache TTL expires.

If you need DLP policy changes to take effect immediately, you can bypass the cache for new requests using the cf-aig-skip-cache header. For more information, refer to Caching.

DLP policies are configured at the gateway level and apply uniformly to all requests passing through that gateway. There is no per-request header to select specific DLP profiles or to bypass DLP scanning for individual requests.

If you need different DLP policies for different use cases (for example, per-tenant policy variance in a multi-tenant application), the recommended approach is to create separate gateways with different DLP configurations and route requests to the appropriate gateway based on your application logic.

AI Gateway DLP uses the same detection profiles as Cloudflare One's DLP solution. Profiles are shared account-level objects, so you can reuse existing predefined or custom profiles across both Gateway HTTP policies and AI Gateway DLP policies.

Key differences from Cloudflare One Gateway DLP:

• No Gateway proxy or TLS decryption required - AI Gateway inspects traffic directly as an AI proxy, so you do not need to set up Gateway HTTP filtering or TLS decryption.
• Separate policy management - DLP policies for AI Gateway are configured per gateway in the AI Gateway dashboard, not in Cloudflare One traffic policies.
• Separate logs - DLP events for AI Gateway appear in AI Gateway logs, not in Cloudflare One HTTP request logs.
• Shared profiles - DLP detection profiles (predefined and custom) are shared across both products. Changes to a profile apply everywhere it is used.

For more information about Cloudflare's DLP capabilities, refer to the Data Loss Prevention documentation.

To enable DLP for your AI Gateway:

1. Set up DLP policies for your AI Gateway
2. Configure detection profiles and response actions
3. Monitor DLP events through the Cloudflare dashboard

• Set up DLP for AI Gateway
• Cloudflare Data Loss Prevention
• AI Gateway Security Features
• DLP Detection Profiles