The story behind the cartoon
Cloudflare has disclosed that 155 responses from its GLM-5.3 model on Workers AI contained data from other customers' requests, between August 27 and September 9. All of them were answers to prompts longer than 681,280 tokens — the size at which the leak triggered.
Two problems overlapped. The model was deployed with a 1,048,576-token context window, but the prefill GPUs' KV cache pool held only 681,280 tokens — a mismatch Cloudflare says it knew about before launch but expected the engine to reject. And the open-source SGLang engine marked an over-length prompt as rejected but did not abort it, so the decode node generated a response from stale cache pages still holding earlier requests' data.
The leak was reported through Cloudflare's HackerOne program on September 8 by Gadi Evron of Knostic. Cloudflare says it reproduced the issue on September 9 and took the model offline within 40 minutes of declaring an incident, restoring it the same day with a corrected configuration and a patched engine — and found no evidence of deliberate exploitation or any other affected Workers AI model.
The joke is the checkmark: the machine did exactly what it was built to do — deliver an answer, stamped SUCCESS — for exactly the wrong person. The green tick isn't lying. It's just not checking whose data it's ticking.
SHUT UP AND COMPUTE.
OPEN DOORS
DECIDES IN 39MS
THE BOUNCER