
Security incident • Cloud AI tenant isolation[1][2]
Cloudflare says 155 GLM-5.3 responses exposed cross-customer data; the incident is resolved[1][3]
Cloudflare says a Workers AI tenant-isolation failure caused 155 responses from its GLM-5.3 model to contain data from other customers’ earlier requests.[1]
Reporting cutoff: 2026-10-26T22:35:49-04:00 (EDT; America/New_York) / 2026-10-27T02:35:49Z (UTC).[1][2][3]
Reader decision
Direct exposure predicate: the confirmed Cloudflare incident was limited to @cf/zai-org/glm-5.3 responses between August 27 and September 9, 2026, and every response Cloudflare identified was tied to a prompt longer than 681,280 tokens.[1]
What happened
Cloudflare configured GLM-5.3 with a 1,048,576-token context window while the prefill KV-cache pool could hold 681,280 tokens. In SGLang’s disaggregated prefill/decode path, an over-limit request could be marked for rejection yet continue through the handoff, allowing decode to read cache pages left by earlier requests.[1][2]
Cloudflare says it reproduced the issue, made the model private on September 9, enlarged the pool, patched the rejection path, validated both changes, and restored the model that evening.[1][2]
Scope and current status
Cloudflare says GLM-5.3 was the only Workers AI model affected and that it found no evidence of deliberate exploitation.[1]
Those are provider findings, not proof that every exposure is recoverable from telemetry or that each affected response was harmless.[1]
The associated Cloudflare Status incident is marked resolved; it began at 14:25 UTC and was resolved at 20:09 UTC on September 9.[3]
Practical guidance for Cloudflare customers
- Check whether the predicate fits. Identify use of the named GLM-5.3 Workers AI model during the August 27–September 9 window, especially prompts above the stated threshold.[1]
- Do not widen the claim. Cloudflare says no other Workers AI model was affected; this disclosure does not establish a platform-wide Cloudflare outage.[1][3]
Guidance for self-hosted SGLang operators
SGLang says v0.5.21 protects Mooncake and NIXL deployments by default, while MORI deployments and generation-tagged acknowledgement handling require a build at or after the fixes included in v0.5.22.[2]
Cloudflare recommends measuring KV-pool capacity against configured context length on every prefill and decode node and upgrading to SGLang v0.5.22 or later.[1]
What remains unknown
The public disclosures do not identify the affected customer accounts, say whether 155 responses map to 155 unique customers, or describe the sensitivity or volume of every exposed fragment.[1][2]
No Florida or Treasure Coast organization was identified in the reviewed first-party disclosures.[1][2]
Primary sources
- Cloudflare: “Cross-request KV cache exposure in the GLM-5.3 Workers AI model, and our remediation with SGLang” — published October 26, 2026.
- SGLang Team: “SGLang security update: improving request isolation” — published October 26, 2026.
- Cloudflare Status: “Cloudflare Workers AI increased errors” — resolved September 9, 2026.
