A bug in SGLang exposed KV cache pages across requests for GLM 5.3 on Workers AI. The issue stemmed from a KV cache pool smaller than the model’s context window. A second bug let rejected long prompts pass a placeholder to decode nodes. Those nodes then read stale cache pages from earlier requests. Cloudflare took the model offline on September 9 after a researcher’s report. It restored service the same day with a larger pool and a patched engine. Only 155 responses were affected, all for prompts over 681,280 tokens. No other Workers AI model was impacted. Cloudflare and SGLang maintainers released fixes in SGLang v0.5.21 and v0.5.22. Operators are urged to update to v0.5.22 or later.
What changed
GLM 5.3 on Workers AI returned only its own responses to each customer.
What it unlocks
Preventing cross-request KV cache leaks in disaggregated SGLang deployments.
- 155 affected responses
- 681,280 tokens KV cache pool limit
- 1,048,576 tokens context window
- Aug 27–Sep 9, 2026 exposure window
What you need to act on it
- using GLM 5.3 on Workers AI
- prompts longer than 681,280 tokens
Sources