On October 26, 2026, Cloudflare disclosed that a bug in one hosted AI model on Workers AI returned pieces of other customers' requests in 155 responses between August 27 and September 9. The fix is in, and Cloudflare says it found no sign of deliberate abuse. The lesson for any business using hosted AI is simple: data isolation is a property of your vendor's setup, so you need to ask about it, test what you can, and limit what you send.
What happened in the Cloudflare Workers AI incident?
According to Cloudflare's post-incident write-up, the problem affected the GLM-5.3 model on Workers AI. Some responses contained data from other customers' requests made shortly before. All 155 affected responses came from very long prompts, longer than 681,280 tokens. Cloudflare says no other Workers AI model was affected.
A security researcher reported the issue through Cloudflare's bug bounty program on September 8. Cloudflare reproduced it, declared an incident on September 9 and took the model offline that afternoon. It published the details on October 26, together with the open source project involved.
Cloudflare also says it "found no evidence of deliberate exploitation." That is good news, but it does not change what the data exposure means for customers whose prompts were caught in it.
Why did one customer's data show up in another's response?
Large language models keep a working memory of the request they are processing, often called the KV cache. Hosted services share hardware across many customers to keep prices down, so that memory is reused quickly from one request to the next.
Two problems lined up. The deployment allowed a context window of 1,048,576 tokens, but the memory pool for the first processing stage held only 681,280 tokens. Separately, a bug in SGLang, the open source serving software, swapped an over-length prompt for a placeholder instead of rejecting it. The next stage then read memory pages that still held another user's data.
Cloudflare's fixes include roughly doubling the memory pool, patching SGLang so oversized requests are rejected before any data moves, and adding deployment checks that compare memory size to the allowed context length.
What does this mean if your business uses hosted AI?
We do not read this as a reason to stop using hosted models. Shared infrastructure is how most businesses get affordable AI. We read it as a reminder that isolation between customers is a configuration detail inside your vendor, not something your contract or your settings guarantee on their own.
Three points stand out:
- The bug sat in widely used open source serving software, so other providers that run it could face similar risks until they patch.
- The trigger was a configuration mismatch that ordinary testing did not catch.
- The exposure lasted about two weeks before an outside researcher found it.
What should you ask your AI vendor about data isolation?
When we help clients choose an AI provider or model host, we ask these questions in writing:
- How do you keep one customer's prompts and outputs separate from another's while requests are processed, not only at rest?
- Which serving software do you run, and how fast do you apply security patches to it?
- Do you check that context limits match the hardware and memory behind each model?
- How will you notify us if our data, or data returned to us, was involved in an incident?
- Do you offer dedicated capacity or a private deployment for sensitive work, and at what cost?
A vendor that answers clearly and quickly is a better partner than one that points only to a certification badge.
What can you control on your side?
You cannot fix your vendor's serving stack, but you can shrink what is at risk. Send the model only the fields it needs. Strip names, account numbers and health details from prompts where the task allows it. Avoid very large, catch-all prompts when a smaller, focused request will do.
Treat responses as untrusted input too. If an output contains a customer name, number or document that your request never included, log it, stop the workflow and report it to the vendor. This is the same principle we apply to task-scoped AI agent permissions: give each step only what it needs.
Finally, build your integration so you can switch models without a rewrite. Our guide to avoiding AI model lock-in covers how to keep that option open.
Should you self-host AI models instead?
Self-hosting gives you your own hardware boundary, but it also makes you the operator. This incident came from serving software and a configuration mismatch, which are exactly the things a self-hosting team would have to manage. For most small and mid-sized businesses, a reputable hosted provider plus careful data minimization is the safer path. Dedicated capacity is worth pricing for workloads that handle regulated or highly sensitive data.
Frequently asked questions
Only requests to the GLM-5.3 model on Workers AI between August 27 and September 9, 2026 were in scope, and only very long prompts. If you used that model then, contact Cloudflare for details.
Cloudflare says no other Workers AI model was affected. Other hosts that run the same open source serving software should confirm they have applied the fixes.
It is the short-term working memory a model uses while it processes a request. On shared hardware, that memory is reused across requests, so it must be cleared and kept separate.
List which AI services receive customer data, send each vendor the isolation questions above, and remove any data from prompts that the task does not need.
Sources
Check your AI vendors before the next incident
Send us a list of the AI services that touch your customer data. We will help you put the isolation questions to each vendor, cut the data your prompts do not need and design integrations that let you switch models without a rewrite. Our API integration and custom development teams can build the safeguards.

