SGLang bug leaked other users' data in long prompts; Cloudflare fixed
Original title：Cross-request KV cache exposure in the GLM-5.3 Workers AI model, and our remediation with SGLang
What happened
On Cloudflare Workers AI, very long prompts to the GLM-5.3 model could return data from other customers; 155 responses were affected. A misconfiguration plus a bug in the open-source SGLang engine caused it, and it is fixed. Operators running SGLang themselves should update.
Who is affected
- Anyone self-hosting SGLang
- Workers AI GLM-5.3 users
What to do
If you run SGLang yourself, update to v0.5.22 or later and verify the KV cache pool is at least the context length. Workers AI users need no action, but review any long prompts sent during the affected window.
Affected products
| Product | Affected | Fixed in |
|---|---|---|
| SGLang | v0.5.20より前(分離prefill/decode構成で影響。正確な範囲は要確認) | v0.5.22以降を推奨(修正はv0.5.20、追加修正はv0.5.21・v0.5.22) |
| Cloudflare Workers AI (@cf/zai-org/glm-5.3) | 2026-08-27〜2026-09-09 | 2026-09-09に修正・復旧済み |
Why this flag
悪用の証拠はなく、Cloudflare側は修正済み。影響は自前でSGLangを分離構成で運用する場合などに限られ、条件付きの更新対応のためYELLOW。
Yellow flag: caution, slow down — Conditional updates, investigation or recovery follow-ups, and regulatory deadlines.
Sources
- Source：Cloudflare Blog ↗