Cloudflare Workers AI: 155 responses leaked other users' data

The leak stemmed from two overlapping problems, according to Cloudflare. GLM-5.3 was published with a context length of 1,048,576 tokens, but the GPU-side KV cache held only 681,280 tokens; Cloudflare knew of the mismatch before launch but assumed the engine would error on oversized prompts. Separately, a bug in the open-source SGLang inference engine, running in a disaggregated setup that splits prompt processing and answer generation across GPUs, meant a rejection flag was not passed to the decode side. The decode side then read leftover memory from a previous request and produced an answer. Cloudflare also noted that KV cache pages are not zeroed per request for performance reasons. The model was taken offline on Sep 9 and republished with a fixed configuration on Sep 10. For operators running SGLang in a disaggregated setup, Cloudflare recommends updating to v0.5.22 or later. The incident highlights that HTTP 200 responses can carry incorrect content, and that client-side checks such as canary markers and pre-send redaction can help detect or prevent data mixing.