← Back to blog
#Ciberseguridad#IA#Nube#Privacidad de datos#Cloudflare

Cloudflare Admits an AI Model Returned Other Customers' Data: What to Learn If Your Business Uses Cloud AI

Servers in a data center enclosure with network cables plugged in

When you buy an AI service in the cloud, you assume that what you type in the chat or send through the API is yours alone. On October 26, 2026, Cloudflare published an incident analysis showing that this assumption can fail: some Workers AI customers received responses containing data from other customers' requests when using the GLM-5.3 model. It is a useful read for any company that already uses AI in its processes.

What happened, according to Cloudflare

According to Cloudflare's own report, the problem affected the @cf/zai-org/glm-5.3 model on Workers AI starting August 27, 2026, the day it launched. A security researcher, Gadi Evron (Knostic), reported it on September 8 through HackerOne. Cloudflare set the model to private on September 9 (requests returned an HTTP 400 error) and restored it the same day, fixed.

The company's numbers are specific: 155 responses were affected, all to requests longer than 681,280 tokens (that is, extremely long inputs, hundreds of thousands of words). Cloudflare states it found no evidence of deliberate exploitation and that no other Workers AI model was affected.

Why it happened: two flaws that added up

The report explains two causes that together opened the door:

  • An inconsistent configuration. The model shipped advertising a context window of 1,048,576 tokens, but the memory that stores intermediate computations (the KV cache) on the "prefill" servers only held 681,280 tokens. It is like selling a 1,000-square-meter warehouse when there is room for only 650.
  • A bug in SGLang, the open-source engine that serves the model. When a request exceeded that space, the engine marked it as rejected but still transferred cache data to the next stage, which ended up "reading" memory pages that belonged to earlier requests from other users. Cloudflare says the bug had already been reported to the SGLang project on July 6, 2026 (issue #30233).

How they fixed it

Cloudflare enlarged the cache to about 2.04 million tokens, roughly twice the advertised context window, so no valid request reaches the faulty path. In the engine, rejected requests are now dropped before any data is transferred. Those fixes reached SGLang in versions v0.5.20 and v0.5.21 (the latter released October 2), and the latest version cited, v0.5.22, came out on October 19. For those who run SGLang themselves, Cloudflare recommends upgrading to v0.5.22 or later and checking that the cache size matches the configured context.

One honest detail from the report: it does not include a list of actions for customers; it only asks that any security finding in Workers AI be reported via HackerOne. What follows in this article are our own recommendations, not Cloudflare's.

What this means for your business

This case does not say "don't use cloud AI." It says that separation between customers is a property of the provider, not something you control, and that it can fail over configuration details. For a small business, the real risk depends on what data you send:

  • Send the AI only what is necessary. If the model does not need the national ID, the customer's tax number or the account number to answer, remove them before sending. What is not sent cannot leak. This also fits what we explained about personal data, chatbots and agents.
  • Be careful with huge contexts. In this incident, only extremely long requests triggered the flaw. Before "pasting the whole file" into a model, ask yourself whether it is needed.
  • Ask your provider. A good AI provider should be able to explain how it isolates its customers, whether it publishes incident reports and how quickly it notifies. Cloudflare's transparency here is exactly the kind of answer to look for.
  • If you host your own models, check the version of your inference engine and compare the available memory with the context you advertise.
  • Define what is sensitive. A simple rule: passwords, API keys, health data, confidential contracts and customer databases do not go to AI services without a written agreement.

For developers

If your product offers AI to several customers on the same GPU or engine, this incident is a case study: test what happens when a request exceeds limits, validate that errors are cut off before sharing state between users, and monitor cache integrity where the engine allows it. Cloudflare mentions corruption-detection options in SGLang that are worth reviewing in the project's documentation. It is also a good time to revisit the risks and uses of AI in cybersecurity.

How we approach it at SimCodec

At SimCodec we design AI solutions for companies with data first in mind: what is sent, where, with what limits and who can see it. That includes voice and WhatsApp AI agents with clear rules and a review of the cloud and network infrastructure they run on. If you want to review what information your company is already sending to AI services, write to us or call our AI assistant Cyntia at +593 99 726 6838.

Sources

← Back to blog Get a quote →