When the inference engine leaks: Cloudflare GLM-5.3 cross-tenant KV cache exposure

On October 26, 2026, Cloudflare and maintainers of the open-source inference engine SGLang jointly disclosed that the GLM-5.3 model on Workers AI experienced cross-request KV cache exposure: between August 27 and September 9, responses to 155 very long prompts mixed in content from other customers' requests. Cloudflare's postmortem states that the prefill GPU's KV cache pool was only 681,280 tokens, smaller than the 1,048,576-token context window; very long prompts were replaced with placeholders but not aborted, and decode performed attention computation on stale pages that still contained other users' state and returned HTTP 200. SGLang's advisory describes a separate mechanism: asynchronous RDMA writes from aborted requests may silently overwrite other users' caches on slots that have been reallocated, affecting disaggregated deployments from v0.4.x to v0.5.21 and leaving no errors or logs.