Cloudflare makes AI Security for Apps generally available

Cloudflare made AI Security for Apps generally available on March 11. The GA announcement describes a security layer for public AI application endpoints, with endpoint discovery, incoming-prompt detection, and policy controls integrated with Cloudflare’s application security platform.

Discovery comes before enforcement

Cloudflare says AI endpoint discovery is available across customer plan tiers, while the broader product is offered to Enterprise customers. Discovery identifies endpoints that appear to carry AI traffic, giving teams an inventory to review before they apply blocking rules. An endpoint classification is evidence for investigation, not proof of ownership or business criticality.

For inspected requests, the product runs each prompt through modules for prompt injection, personally identifiable information exposure, and sensitive or toxic topics. Detection results become metadata for custom WAF rules. The GA release also adds custom topics, letting operators define subjects and receive a relevance score that they can use to log, block, or otherwise handle the prompt.

Define the failure path

A rollout should begin in observation mode with a representative sample of incoming prompts and request-body formats. Teams need to measure false positives by route and use case, especially where ordinary business fields overlap with restricted topics. They should also decide whether a detection blocks a request, records an event, or triggers a custom response or separate review path.

Because inspection occurs before prompts reach the application infrastructure, server-side authorization and data minimization remain separate controls. A prompt filter cannot determine whether a caller should see a customer record. The strongest deployment combines endpoint inventory, narrowly scoped policies, and application logs that let responders reconstruct what was allowed or denied without retaining more prompt content than the organization needs.