Summary
A SGLang misconfiguration and disaggregated-prefill bug caused 155 GLM-5.3 responses (prompts >681k tokens) to leak other tenants’ KV cache data between Aug 27–Sep 9; Cloudflare remediated with pool sizing and engine patches coordinated with SGLang.
Source Analysis
Research Notes
Additional sources
- Cloudflare + SGLang coordinated blog posts (Oct 26)
- Hacker News discussion (Gadi Evron / Knostic reporter)
- DataStudios, dev48 summaries
Verified
- 155 affected responses; prompts >681,280 tokens; window Aug 27–Sep 9, 2026
- Root cause: KV pool < 1,048,576 context + SGLang rejection-path bug in disaggregated prefill/decode
- Remediation: pool ~2.04M tokens, engine patch, model restored Sep 9
Context
- Teachable tenant-isolation failure for any multi-tenant LLM host, not only Cloudflare
- Links to prior glm-5-3 security discourse (anthropic-glm-5-3-cyber-capabilities)
Related wiki
Draft Article
PreScreening Notes
- Domain fit: ai, software — Workers AI tenant isolation incident disclosure
- Newsworthiness: KV cache cross-tenant leak (155 responses); SGLang/config root cause and remediation — actionable for AI infra teams
- Audience: High — Cloudflare Workers AI users and platform engineers
- Duplicates: None
- Recency: Blog publication 2026-10-26T22:41:48Z (~7.4h); incident dates older but disclosure is the news event
Evaluation Report
News value: Disclosure is fresh; incident window Aug–Sep but post-mortem is the event. Impact: anyone running multi-tenant LLM inference (not only Cloudflare). Prominence: Cloudflare + SGLang stack. Proximity: Turkish teams on Workers AI or similar hosted inference need mitigation checklist. Novelty: KV cache cross-request leakage is a distinct, teachable failure mode.
Audience fit: High for software engineers and AI infra; actionable remediation narrative.
Risk & ethics: Authoritative vendor post-mortem — cross-check scope (155 responses, token threshold) against blog; no sensationalism.
Publication strategy: standard — timeline, root cause, tenant isolation lessons; priority high.
Suggested Angle
“Cloudflare Workers AI’da KV cache sızıntısı: 681k token eşiği ve SGLang kök nedeni” — Multi-tenant inference’da yapılandırma ve disaggregated prefill riskleri; operasyonel kontrol listesi.
Editorial Notes
Onay: standard — teknik post-mortem + tenant isolation dersleri.
Reporting talimatları: 155 yanıt / >681k token / Aug 27–Sep 9 penceresini timeline ile ver. SGLang + pool sizing düzeltmesini kök neden olarak netleştir. Son paragrafta Cloudflare’a özel olmayan tenant-isolation checklist (2–4 madde).
Başlık önerileri:
- Cloudflare, GLM-5.3’te KV cache sızıntısını açıkladı: 155 etkilenen yanıt
- Multi-tenant LLM riski: SGLang yapılandırması ve 681 bin token eşiği
Zorunlu noktalar: Etki ölçeği; disaggregated prefill bug; remediation (pool ~2,04M, patch); model 9 Eylül’de geri.
Published article: published/2026-10-26-cloudflare-glm-kv-cache-exposure.md