The company says a configuration mismatch triggered a bug in SGLang, the inference engine used for the deployment.
According to Cloudflare, the bug could expose KV-cache pages across requests. KV cache stores intermediate model state used during inference; cross-request exposure therefore concerns separation of request context within the managed service. Cloudflare describes the post as a remediation undertaken with SGLang.
The report identifies the affected model deployment as GLM-5.3 on Workers AI and attributes the issue to the combination of a configuration mismatch and an SGLang bug. It does not, from the supplied source material, establish which customers or requests were affected, whether any exposed cache content was accessed, what data it contained, or the duration and scale of the exposure.