GLM-5.3 Workers AI KV Cache Leak Fixed

Between August 27 and September 9, 155 long-prompt responses from the GLM-5.3 Workers AI model unintentionally included data from other customers. A security researcher, Gadi Evron of Knostic, reported the issue on September 8 through Cloudflare's HackerOne program. Cloudflare reproduced it the next morning and shut the model down in 40 minutes. The service was back online on September 9 after applying a configuration change and engine patch. The first issue was a mis-matched KV cache pool of 681,280 tokens against a 1,048,576-token context window. The second was a SGLang bug that let an over-length request generate a placeholder. It then transferred the placeholder's KV cache, leaving stale pages from earlier users intact. To fix it, Cloudflare expanded the KV pool to about 2.04 million tokens, enough for the full 1-million-token context. They also patched SGLang so rejected requests are retired with an error before any KV transfer occurs. Cloudflare says no evidence of deliberate exploitation was found and only ordinary long-context usage triggered the leak.