[UPDATE] Cloudflare discloses cross-tenant KV-cache exposure in Workers AI GLM-5.3: 155 responses carried others' data
Cloudflare disclosed that between Aug 27 and Sep 9, 155 Workers AI responses from GLM-5.3 contained fragments of other customers' requests. The cause was a context window larger than the prefill KV pool, combined with an SGLang disaggregated-serving rejection path that passed stale cache pages to decode. Cloudflare says no other model was affected and it found no evidence of deliberate exploitation. In a coordinated post, the SGLang maintainers describe upstream fixes and a related race condition, and recommend operators run v0.5.21/v0.5.22 or later. The incident raises tenant-isolation questions for shared inference platforms; watch for further disclosures from operators running disaggregated serving.