BreakingCyber Security· 🧪 Worth Testing

Critical KV cache isolation vulnerability in SGLang impacts multi-tenant AI deployments

In short: Cloudflare identified a security flaw in the SGLang inference engine that caused cross-request data exposure for the GLM-5.3 model. Requests exceeding the configured KV cache pool size were incorrectly handled, causing the decode node to generate responses using stale KV cache pages belonging to other users.

Source: Cloudflare Security27 Oct 2026, 00:01CloudflareGLM-5.3 Original article ↗

This summary was generated automatically by AI from Cloudflare Security's publication. It is our own text, not a copy of the original — facts, figures and quotes belong to the source, linked above and below.

What changed?

1Updated SGLang to v0.5.22 to fix admission queue rejection logic
2Corrected KV cache pool configuration to accommodate full context windows
3Implemented pre-deployment validation of KV pool size against context length
4Added checksum validation for transferred KV pages

Why it matters

This incident highlights the severe risks of memory management in high-performance inference engines. For teams building multi-tenant AI systems, it demonstrates that logic errors in request rejection paths can bypass tenant isolation, leaking sensitive session data across users.

What it means for AI agents and contact centers

If your infrastructure uses disaggregated SGLang deployments, ensure you are running v0.5.22+. During testing, always validate that your KV cache pool size exceeds your model's maximum context length to prevent silent failures that lead to memory safety issues.

🧪 Worth Testing

It is essential to audit SGLang deployments and memory management practices to ensure tenant isolation remains robust during error states and high-load scenarios.

SGLang· New

KV Cache Tenant Isolation SGLang Security Workers AI CVE

Sources

Cloudflare SecurityOfficialPrimary source
„Cross-request KV cache exposure in the GLM-5.3 Workers AI model, and our remediation with SGLang“
27 Oct 2026, 00:01
Licence: robots.txt Content-Signal: ai-input=yes · our summary (content changed)
Original article →

Published by source: 27 Oct 2026, 00:01
Found by our system: 27 Oct 2026, 00:21
Summary generated: 27 Oct 2026, 00:21

This article was written by AI from the original source. Facts, numbers and prices come from the source; missing values are marked “Not specified”. Legal notice, copyright and privacy