Cloudflare fixes KV cache leak in GLM-5.3 AI model

Cloudflare's Workers AI platform discovered a configuration mismatch in the GLM-5.3 language model that triggered a bug in the underlying SGLang inference framework. The bug caused KV (key-value) cache pages—data structures that store intermediate computation results—to leak across separate user requests, potentially exposing sensitive information processed by the model.

The issue stemmed from how the model was configured relative to how SGLang handles cache isolation. Cloudflare identified the problem, understood its root cause, and deployed a fix to prevent cache pages from persisting or crossing request boundaries.

This type of vulnerability highlights the complexity of safely running large language models at scale and the importance of careful configuration and testing in production inference systems.