Latest update
Prompts over 681K tokens got other customers' data: Cloudflare discloses GLM-5.3 leak in 155 responses
First seen on X 13 hours ago@newl0re ♥ 1511 views
"They knew about the gap. Long-context tests came back HTTP 200, so it looked fine. 155 responses ended up carrying other customers' data," developer @newl0re posted on Oct 27 (151 likes).
Cloudflare said in a blog post on Oct 26 that between Aug 27 and Sep 9, a small number of customers using the GLM-5.3 model on its Workers AI inference service received responses containing data from other customers' requests. In all, 155 responses were affected, every one to a prompt longer than 681,280 tokens.
Two problems overlapped. Cloudflare launched GLM-5.3 with a 1,048,576-token context window, but the KV cache pool on its prefill GPUs only held 681,280 tokens. The company knew about the mismatch before launch but expected the engine to return an error, and its long-context tests came back successful. Then a bug in the open-source inference engine SGLang kicked in: over-length prompts were swapped for a one-token placeholder, and the decode GPUs generated answers while attending over stale cache pages left by earlier requests.
Security researcher Gadi Evron of Knostic reported the issue through HackerOne on Sep 8. Cloudflare reproduced it the next morning, pulled the model within 40 minutes of declaring an incident and restored it the same day with a fix. It said it found no evidence of deliberate exploitation and that no other hosted model was affected. The disclosure was coordinated with SGLang.
“Cloudflare shipped GLM-5.3 with a 1,048,576-token context on a KV cache that held 681,280. They knew about the gap.”
