155 answers from someone else's cache
Cloudflare says a configuration error and a bug in the SGLang serving engine let its GLM-5.3 model answer 155 over-long prompts using memory left by other customers' requests between 27 August and 9 September. It found no evidence of deliberate exploitation.
HaeB, CC BY-SA 4.0, via Wikimedia CommonsCloudflare said on Monday that 155 responses from the GLM-5.3 model on its Workers AI platform drew on data left in memory by other customers' requests between 27 August and 9 September. A security researcher reported the flaw through its bug bounty programme.
The company's write-up, by Michelle Chen and published as a coordinated disclosure, says "In total, 155 responses were affected, all to prompts longer than 681,280 tokens." Those responses "contained data from other customers' requests", including fragments of other users' prompts and partial answers, and reached what it calls a small number of customers using the model. The report came from Gadi Evron of Knostic through HackerOne. "We found no evidence of deliberate exploitation," Cloudflare said, adding that the affected requests looked like ordinary long-context use. The account is Cloudflare's own and has not been independently verified.
The mechanism is a mismatch between two numbers. Large language models keep a key-value (KV) cache, a working memory of the prompt they have already processed, so they do not have to recompute it for every new word. GLM-5.3 accepts prompts of up to 1,048,576 tokens, but the pool of cache memory set aside for processing prompts held only 681,280. Cloudflare says the mismatch was known before launch, and that pre-launch tests checked for successful HTTP status codes rather than for coherent output, which it calls a critical gap in its testing.
On its own the misconfiguration should have produced an error. Instead, a bug in SGLang, the open-source serving engine Cloudflare uses, swapped over-length prompts for a one-token placeholder in its split prefill-and-decode mode. The queues that admit requests checked whether a request had finished, not whether it had been rejected, so the placeholder went through, and the decoding stage read stale cache pages left by earlier requests and returned a normal HTTP 200 answer. "Every request over the threshold behaved this way," the company wrote.
| Time (UTC) | Event | Elapsed, by our count |
|---|---|---|
| 27 Aug | GLM-5.3 launches on Workers AI with an undersized prefill cache | Exposure begins |
| 8 Sep, 23:11 | Researcher's report arrives through HackerOne | 12 days after launch |
| 9 Sep, 13:22 | Security incident declared; leak reproduced within 25 minutes | 14h 11m after the report |
| 9 Sep, 14:00 | Model set to private; requests return HTTP 400 | 14h 49m |
| 9 Sep, 14:52 | All prefill nodes moved to a corrected configuration | 15h 41m |
| 9 Sep, 19:40 | Model restored to public on a patched build | 20h 29m |
| 2 Oct, 19 Oct | SGLang v0.5.21 and v0.5.22 released | |
| 26 Oct | Coordinated publication | 47 days after the fix |
The fix came in two parts. Cloudflare enlarged the prompt cache pool to about 2.04 million tokens, roughly 1.95 times the context window by our arithmetic, so valid requests never reach the rejection path, and patched the engine so that rejected requests are retired at the admission queue with their original error before any cache is handed over. The upstream change is visible on GitHub as SGLang pull request 38935, merged on 16 September. "Nothing forces the two to agree, and it is our job to check that they do," Cloudflare wrote of the cache size and the context length, which SGLang sets separately.
The numbers give the shape of the exposure. The cache held 65 per cent of the context window, by our arithmetic, so only prompts in the band between 681,280 and 1,048,576 tokens, a span of 367,296 tokens, were affected, which is why the count is 155 responses rather than every request. The model was taken private 14 hours and 49 minutes after the report arrived and was back online on a patched build 20 hours and 29 minutes after it. The write-up does not say whether affected customers were told directly.
Why it matters: inference providers sell long context windows as a feature, and the longest prompts are the ones most likely to carry whole documents or codebases. A shared serving stack that runs many customers through the same memory relies on configuration and admission checks to keep them apart; here both failed at once, and the system reported success. It is the second platform in a week to tell customers that their data may have been reachable by others: Railway asked every customer to rotate its keys after an unauthorised party may have decrypted some environment variables.
Cloud incidents have run through the autumn. In October a ransomware attack took four zones of a SoftBank unit's cloud offline, as Komposite reported, and Google said a flaw in BeyondTrust's remote support tool had been exploited from late August. Cloudflare's case differs in one respect: no attacker was needed.