Cross-request KV cache exposure in the GLM-5.3 Workers AI model, and our remediation with SGLang
Cloudflare knew before launch that the cache was a third smaller than the context window, and its long-context tests passed because they checked the status code and never read the reply.