Cloudflare: GLM-5.3 Leaked Other Users' Chats on Workers AI

Cloudflare disclosed a data leak on its Workers AI cloud, a service that runs AI models for other companies' applications. GLM-5.3, a model built by China's Z.ai for coding agent tasks, would sometimes insert a chunk of another customer's chat into its response. This only happened on very long requests, above 681,280 tokens (tokens are the units a model breaks text into). Over two weeks, from August 27 to September 9, this affected 155 responses in total. A researcher found the problem on September 8. Cloudflare confirmed it within 25 minutes, took the model offline, fixed the issue, and brought it back the same day, September 9.

The cause was not GLM itself. GLM-5.2, which runs on a different Cloudflare setup, was not affected at all. The leak came from Cloudflare's own hosting configuration — a memory pool that turned out too small for the model's load — combined with a bug in SGLang, the open source engine (software that runs the model and handles requests) Cloudflare uses to serve it. For businesses building agents on shared cloud infrastructure with access to other customers' data, the lesson is about hosting and internal plumbing, not about which model answers. Cloudflare confirming the researcher's report in 25 minutes and fixing it the same day is close to what that response should look like.