Cloudflare’s coordinated disclosure confirms cross-request KV-cache exposure on GLM-5.3 Workers AI and documents remediation and operator upgrade requirements.
Cloudflare discloses cross-request KV cache exposure in Workers AI
2026-10-27Verified announcementADOPT
Cloudflare disclosed that 155 GLM-5.3 Workers AI responses exposed data from other customers between August 27 and September 9. A KV pool smaller than the advertised context window combined with an SGLang disaggregated-inference bug, causing over-length requests to attend to stale cache pages. Cloudflare restored the model after correcting the configuration and patching the engine, and reports no evidence of deliberate exploitation or impact to other models.
What this means for you
For teams running disaggregated inference, a successful HTTP response does not prove tenant isolation or coherent long-context behavior. Operators should upgrade SGLang, assert that KV capacity covers the advertised context and test response content across independent tenants.
Evidence & related material
Technical details & scores
| Affected Service | Cloudflare Workers AI @cf/zai-org/glm-5.3 |
|---|---|
| Exposure Window | 2026-08-27 to 2026-09-09 |
| Affected Responses | 155 |
| Trigger | Prompts longer than 681,280 tokens against a 1,048,576-token advertised context |
| Root Cause | KV pool configuration mismatch plus an SGLang rejected-request bug in disaggregated prefill/decode mode |
| Remediation | Cloudflare enlarged the pool, patched rejection handling and hardened production deployments; operators are advised to use SGLang v0.5.22 or later |
| Evidence Limit | Scope and non-exploitation findings are Cloudflare incident-response conclusions |
| Time Basis | Cloudflare incident report published 2026-10-26T22:01:23.889Z; underlying exposure occurred August 27–September 9 |
Technical impact 5
Engineering value 5
Adoption signal 4
Maturity 5
Verification cost 2
Risk 5