Cloudflare GLM-5.3 cross-customer KV-cache exposure postmortem

Evidence record

Cloudflare production postmortem for 155 GLM-5.3 responses contaminated with other tenants’ KV state during August 27 to September 9. It documents context/cache mismatch, admission failure, temporary disablement and reported remediation.

Read the original source ↗

Claim-level citations (4)

• supportsCloudflare inference rejection path leaked another tenant’s cached content: Cloudflare reports 155 AI responses containing content from another customer during an incident window from August 27 through September 9, 2026.

  Incident postmortem; Impact and timeline

• supportsCloudflare inference rejection path leaked another tenant’s cached content: The documented serving mismatch was a 1,048,576-token advertised context against an approximately 681,280-token effective KV-cache limit, with a rejected prefill entering a decode path that returned HTTP 200.

  Technical root cause; Context length and prefill/decode handling

• supportsCloudflare inference rejection path leaked another tenant’s cached content: Cloudflare says it disabled GLM-5.3, enlarged the cache pool to approximately 2.04 million tokens, and fixed rejection handling before restoring access on September 9; it found no evidence of deliberate exploitation.

  Immediate remediation and incident impact, October 26 postmortem

• supportsSGLang: Cloudflare reports that its GLM-5.3 cross-customer incident came from an over-capacity prefill request entering decode, while the SGLang team separately reproduces an asynchronous RDMA cancellation/write-after-reassignment hazard. The two mechanisms must not be counted as the same confirmed production incident.

  Technical root cause and immediate remediation; affected response count