Cloudflare discloses the GLM-5.3 Workers AI cross-request KV cache leak incident and details of the fix
Cloudflare's October 26, 2026, joint disclosure report explains that the GLM-5.3 model on Workers AI encountered issues where prefill KV cache pools (681,280 tokens) were smaller than the context window (1,048,576 tokens). Combined with rejection path defects in SGLang's split inference architecture, this caused extremely long requests to read KV cache data from other users' previous requests.
Return to original text
A short introduction is retained here; the full text from the source website is not reposted. For specific wording and subsequent updates, please refer to the original article.
Open AIHOT original