Cloudflare’s October 26 incident report says its Workers AI model @cf/zai-org/glm-5.3 returned other customers’ request data in 155 responses between August 27 and September 9, 2026. Every affected response followed a prompt longer than 681,280 tokens, the small text units models process.
Long prompts could read earlier requests
The model’s advertised context window, the material it could process at once, exceeded its working-memory pool, called the KV cache. A bug in SGLang, the software serving the model, let oversized requests continue and read memory left by earlier requests.
The length threshold describes requests that received leaked data. It does not establish that data submitted in shorter requests was protected from appearing in someone else’s response.
The service was restored on September 9
Cloudflare says it enlarged the pool, patched the engine, and restored the model on September 9. It found no evidence of deliberate exploitation and says no other Workers AI model was affected.
If your app used this model during that period, review what sensitive material you sent. Base that review on the model and dates, rather than assuming a short prompt ruled out exposure.
For builders who run SGLang themselves, Cloudflare recommends v0.5.22 or later and checking that the working-memory pool can hold the advertised context length.
End of article