Cloudflare AI leaked user data in 155 replies

Cloudflare said on the 26th of October 2026 that its Workers AI service leaked user data in 155 replies. Some customers got answers holding parts of other customers' requests.

Workers AI lets businesses use AI models on Cloudflare's servers. The problem hit only the GLM 5.3 model, from the Chinese firm Z.ai, between the 27th of August and the 9th of September.

What went wrong

The model was set to take prompts of up to 1,048,576 tokens, the small pieces of text AI reads. The memory on the first stage of the servers held only 681,280 tokens, Cloudflare wrote.

A bug in SGLang, the free software Cloudflare uses to run models, let longer prompts run on memory still holding earlier users' data. All 155 replies came from prompts over that size.

How it was found and fixed

Researcher Gadi Evron of the security firm Knostic reported the problem on the 8th of September. Cloudflare took the model offline the next day and brought it back fixed that evening.

Cloudflare said it found no sign that anyone used the bug on purpose. It asked other SGLang users to update to version 0.5.22 or later.

Below we break down the numbers

Before the fix, the memory held about 65 percent of the prompt size the model promised, by our arithmetic. After the fix it holds about twice the promised size.

What is confirmed and what is not

Cloudflare said a small number of customers were affected but gave no count. It did not say whether those customers were told.

Key facts

The problem hit only the GLM 5.3 model on Workers AI.

It ran from the 27th of August to the 9th of September 2026.

Gadi Evron of Knostic reported it on the 8th of September.

Cloudflare found no sign the bug was used on purpose.