A configuration mismatch in the GLM-5.3 model on Workers AI triggered an SGLang bug that exposed KV cache pages across different requests. The issue has since been resolved by the engineering team to prevent further cross-request data leakage.