Cross-request KV cache exposure in the GLM-5.3 Workers AI model (155 responses)

1. #0110.26
   Initial2026-10-26 22:41 · publication time

   Cross-request KV cache exposure in the GLM-5.3 Workers AI model (155 responses)

   Cloudflare published a root-cause analysis of a Workers AI bug in which GLM-5.3's cross-request KV cache was not properly isolated between users, allowing potential overlap of session memory, traced to a prefill cache capacity mismatch and remediated by pulling the model from serving.

   Source evidence: wondersky

2. #0210.27
   Remediation2026-10-27 10:12 · publication time

   Cloudflare AI用户隔离失效：GPU缓存不够竟照常回答，155条回复泄露其他客户内容

   Disclosed a separate SGLang cache-reuse bug, its fix version v0.5.22, and that the actual leak was fixed on September 9, with no evidence of malicious exploitation and other Workers AI models unaffected.

   State after update: The Cloudflare Workers AI GLM-5.3 user-isolation leak has been fixed; related SGLang cache-management defects were patched with a recommendation to upgrade to v0.5.22; no malicious exploitation was found and other Workers AI models were unaffected.

   Source evidence: theblockbeats

All events · Back to the feed