Cloudflare discloses cross-request KV cache leak in the GLM-5.3 model on Workers AI and coordinated fix with SGLang
Cross-request KV cache exposure in the GLM-5.3 Workers AI model, and our remediation with SGLang
AI summary
Cloudflare 披露,其 Workers AI 平台上 Z.ai 的 GLM-5.3 模型曾发生跨请求 KV cache 泄露,2026 年 8 月 27 日至 9 月 9 日间共 155 条回复包含其他用户请求的数据。受影响回复全部来自超过 681,280 token 的长提示,无证据显示漏洞被刻意利用,其他 Workers AI 模型未受影响。根因是两个问题叠加:prefill 节点的 KV cache 池仅 681,280 token,小于模型 1,048,576 token 的上下文窗口,而推理引擎 SGLang 在分离式 prefill/decode 模式下会把被拒的超长请求替换为单 token 占位符后照常传输,decode 节点因此读取了先前请求遗留的缓存页。
Why it was selected
The article provides a method for checking KV pool and context length and upgrade recommendations, so operators of other SGLang disaggregated deployments can directly check for similar cross-request leak risks.
Read original