Cloudflare discloses cross-customer data leak in GLM-5.3 caused by an SGLang bug: 155 responses affected
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
A bug in an open-source inference engine under a rare configuration led to data leaking between customers of a cloud model.
On October 26 Cloudflare published a report on a customer isolation incident in the GLM-5.3 model on Workers AI. From August 27 to September 9 a small number of customers received responses containing data from other customers' requests to the same model; 155 responses were affected, all to prompts longer than 681,280 tokens. According to Cloudflare, the cause was a configuration mismatch (the model advertised a 1,048,576-token window while the GPU memory pool held 681,280) that triggered a bug in the SGLang engine: the decode node read stale KV cache pages. Security researcher Gadi Evron of Knostic reported the vulnerability on September 8; on September 9 Cloudflare restored the model with a corrected configuration and an updated SGLang, and other Workers AI models were not affected. Cloudflare says it found no evidence of deliberate exploitation. The report was released together with a post from SGLang as a coordinated disclosure.
Sources 1