Skip to main content

Edge Guardrails Pilot on Cloudflare

Local mirror of Confluence page Edge Guardrails Pilot on Cloudflare.

Confluence page ID: 5548573323 Parent folder ID: 6018662571 Remote version: 4 Last remote update: 2026-03-03T15:18:43.668Z

Context​

Today, several guardrail capabilities run in centralized cloud environments, such as AWS-based services and external model endpoints. This can add latency, cross-provider dependencies, and operational complexity when applying consistent guardrail behavior across multiple ADC solutions.

Cloudflare proposes an edge-native AI execution model where inference and application logic run close to users and systems.

Key proposition:

  • Run AI inference at the edge with low-latency execution.
  • Use serverless AI APIs that are OpenAI-compatible for easier integration.
  • Leverage Cloudflare performance work on efficient inference engine design.
  • Keep workloads inside one edge platform to reduce data movement and simplify architecture.

The Problem We Want to Solve​

The current guardrail approach can be fragmented across providers and runtime locations.

Key pain points:

  • Higher end-to-end latency when guardrail checks are remote from the calling application.
  • Operational overhead from stitching together multiple cloud/model services.
  • Limited portability of guardrail behavior across products.
  • Reduced control over where and how guardrail logic executes.

Goal​

Validate whether core guardrail capabilities can move to Cloudflare Edge and operate as a shared internal platform service for multiple ADC solutions.

By the end of the pilot, answer:

  • Can the guardrail pipeline run reliably on Cloudflare Edge with acceptable latency and throughput?
  • Can dependency on external guardrail infrastructure be reduced while preserving quality and safety coverage?
  • Can one edge-native guardrail service be reused across multiple ADC use cases with minimal integration effort?
  • What are the cost, observability, and operational tradeoffs versus current cloud-hosted patterns?

What We Are Proposing​

Build a pilot guardrail system that runs directly on Cloudflare Edge with Workers and Workers AI-compatible integration patterns, instead of relying primarily on AWS-hosted guardrail components or externally orchestrated model services.

Pilot proposal:

  • Create an edge-native guardrail gateway that receives model requests from ADC applications.
  • Execute policy checks at the edge, including input/output moderation, prompt policy enforcement, safety filters, and rule-based controls.
  • Support pluggable model backends while keeping guardrail orchestration inside Cloudflare.
  • Expose a unified API contract so multiple ADC products can adopt the same guardrail layer.
  • Instrument latency, policy hit rates, blocked/allowed outcomes, and operational cost for comparison with baseline architecture.

Expected value:

  • Lower latency for guardrail decisions.
  • Improved architectural control and consistency across solutions.
  • Faster rollout of shared policy updates.
  • A concrete decision basis for broader migration beyond pilot scope.

References​