Cloudflare said its GLM 5.3 model on Workers AI, its hosted AI service, showed some customers other customers' prompts and partly written answers. This ran from August 27 to September 9, 2026, from launch. Prompts over 681,280 tokens triggered it, and 155 replies were hit. Researcher Gadi Evron of Knostic reported it on September 8. Cloudflare restored the fixed model on September 9.
Cloudflare set the model to accept 1,048,576 tokens, but its memory pool held only 681,280. It knew before launch and expected the engine to return an error. Its tests got success codes, so it saw no problem. In SGLang, the open-source engine, an oversized prompt was swapped for a one-token stand-in. The sending machine then said the transfer was done, and the receiving machine read memory pages that are not wiped between requests. Users got a normal reply with no error. Cloudflare found no sign of deliberate abuse. It does not say what the prompts contained.
A success code does not show that customers are kept apart. Cloudflare's tests checked status codes and speed, not whether answers made sense. Buyers of hosted models can ask vendors how they check memory size against prompt size, and whether their tests read the replies. SGLang users should update to v0.5.22 or later.