Cloudflare says 155 GLM 5.3 responses on Workers AI contained data from other requests between August 27 and September 9. A roughly 681,000-token KV cache was smaller than the configured 1,048,576-token context window, and rejected long prompts incorrectly continued through SGLang’s disaggregated serving path. Cloudflare corrected the configuration and engine on September 9, reporting no other affected models or evidence of deliberate exploitation. SGLang’s companion disclosure also examines memory transfers that outlive cancellation, showing why successful HTTP responses alone are insufficient evidence of output isolation.