AI data security · News & analysis
Your AI gateway is also a data system
Cloudflare’s AI Gateway update makes request visibility easier. What teams should decide about prompt retention, access and provider routing.
On August 7, Cloudflare described the convergence of Workers AI and AI Gateway, including a shared request path and easier access to request logging and usage visibility. The announcement also outlines future routing capabilities, distinguishing those plans from features available at publication.
Better observability can help a team investigate failures and understand its AI usage. It also creates a data-handling decision: prompts and responses may contain information that should not be available to everyone who can view an operations dashboard.
Decide what the log is for
Our recommendation is to define the purpose of an AI request log before choosing what to retain. Debugging latency may require timing and model identifiers. Investigating an incorrect answer may require selected input and output. Those purposes do not automatically justify retaining every customer document in full.
Consider an assistant that summarizes a support case. The case may include personal information, confidential attachments or a credential pasted by mistake. Routing the request through a gateway can create another copy of that material. Include the gateway in the data map and access review rather than treating it as invisible plumbing.
Separate operational access from content access
An engineer investigating error rates does not necessarily need to read all prompt bodies. Where your tools allow it, separate the ability to inspect operational metrics from the ability to retrieve sensitive content. Give exceptional access a reason and an expiry, and retain a record of that access.
Review how logging behaves in development, staging and production. Test fixtures can be deliberately synthetic, while production records need a retention decision and a way to remove data when appropriate. Document any redaction before storage, including what it cannot reliably detect.
- Identify which prompts, responses and metadata are stored.
- Assign an owner for retention and access settings.
- Check whether debugging exports create additional copies.
- Review provider destinations before enabling automatic routing or fallback.
Treat a routing change as a data-flow change
A fallback model can improve availability, but changing the receiving provider may also change who processes the data. Keep the allowed destinations explicit. A model name alone may not identify the organization receiving a request, the region involved or the retention arrangement.
The useful outcome is a request path that the engineering and security teams can explain together. You should be able to trace an example request, identify where content was retained and show who could read it. That makes the gateway useful evidence for operating the application, rather than another unreviewed source of sensitive data.
Put your security work in motion.
Book a meeting