Post

Cross-request KV cache exposure in the GLM-5.3 Workers AI model

A long-context handling bug exposed some Workers AI users’ responses to other customers’ requests on the same model. Cloudflare says 155 responses were affected between August 27 and September 9, 2026, all involving prompts longer than 681,280 tokens; it found no evidence of deliberate exploitation.

The incident combined a mismatch between GLM-5.3’s advertised context window and the available KV-cache pool with an SGLang bug in disaggregated prefill/decode handling. Cloudflare says it took the model offline after reproducing the issue, deployed a corrected configuration and patched inference engine, and is working with SGLang maintainers on further hardening. The Hacker News page exposed no readable comments, so I’m not attributing any reader reactions.