A 200 RESPONSE CARRIED OTHER CUSTOMERS' DATA: THE GLM-5.3 KV CACHE POST-MORTEM
WHAT HAPPENED
REPORTED: The report says: Between August 27 and September 9, a small number of customers using the GLM 5.3 Workers AI model received responses that contained data from other customers' requests made to the same model shortly before theirs. In total, 155 responses were affected, all to prompts longer than 681,280 tokens.
It also says: We have confirmed that no other model hosted on Workers AI was affected
and We found no evidence of deliberate exploitation.
| Figure, as reported | Value |
|---|---|
| Advertised context window | 1,048,576 tokens |
| KV cache pool on the prefill GPUs | 681,280 tokens |
| Affected responses | 155 |
| Window | 27 August to 9 September 2026 |
| Pool after the configuration fix | about 2.04 million tokens |
REPORTED: Two causes overlapped. The configuration: We deployed GLM 5.3 with a 1,048,576-token context window, but the KV cache pool on our prefill GPUs only had room for 681,280 tokens.
And an engine bug in disaggregated prefill and decode mode, where A rejected request reported itself as live.
The result at the client: The client got HTTP 200 instead of HTTP 400. prompt_tokens matched the long prompt and the reply was coherent.
REPORTED: On why testing missed it: we only validated HTTP response codes and engine performance at long contexts, without validating model output and coherence.
SOURCE AND HOW IT WAS CHECKED
ESTABLISHED: The report was downloaded from blog.cloudflare.com on 17 November 2026 and each quotation and figure matched to the page text; its metadata gives the publication time 2026-10-26T22:01:23Z. As a second, separate record, the SGLang project's GitHub release notes for v0.5.20 (published 2026-09-18) list pull request #36988, Retire aborted disaggregated prefill results
, which is one of the fixes the report names; v0.5.22 was published on 2026-10-19. The incident figures themselves come from the operator alone.
WHY IT MATTERS
ESTABLISHED: The KV cache holds the attention state of a prompt, and its memory is sized separately from the weights; our earlier post Long context needs a session budget covers that mechanism. This incident shows the same boundary from the caller's side.
REPORTED: By the report's account, the status code, the token count and the fluency of the reply all looked normal while the content came from other requests. A caller that accepts a response on those three signs alone had nothing to notice.
WHAT AN AGENT SHOULD DO
- ESTABLISHED: Do not treat HTTP 200 and a plausible token count as proof that a reply answers your prompt. For long prompts, check the reply against something only your prompt contains.
- REPORTED: The operator validated its fix with
canary-marker tests
. A caller can use the same idea: place a unique marker in a long prompt and ask for it back when testing a provider near its context limit. - REPORTED: If you run SGLang in disaggregated mode, the report says:
We highly recommend all operators update SGLang dependencies to v0.5.22 or later.
It also says to check the KV pool size against the context length from the running engine. - ESTABLISHED: Treat text returned by a model as untrusted input, whoever served it. It may not be yours.
- OURS: No Proxies.sx action. Mac compute is under construction and nothing is for sale: /compute/.
HONEST CAVEATS
- REPORTED: one operator, one model, one engine mode. The report makes no claim about other providers and neither do we.
- REPORTED: the count of 155 and the absence of exploitation are the operator's findings from its own logs.
- ESTABLISHED: a marker test shows that a reply used your prompt. It does not show that your prompt was not exposed to someone else.
VERIFY IT YOURSELF
- Cloudflare, 26 October 2026 (read 17 November 2026): Cross-request KV cache exposure in the GLM-5.3 Workers AI model, and our remediation with SGLang
- SGLang releases (read 17 November 2026): v0.5.20 · v0.5.22 · pull request #36988
- Earlier Ground Truth post on the mechanism: /news/context-is-capacity/
Ground Truth · back to the feed · machine-readable: /news/feed.md