Futureweb

Distributed Cloud

The GLM pool report is October

Cloudflare says 155 GLM-5.3 responses included another customer text. Other Workers AI models were spared. The model was restored on September 9.

The short version

No. The October 26 report covers @cf/zai-org/glm-5.3 only. 155 responses to prompts longer than 681,280 tokens were affected, between August 27 and September 9.

The report

On October 26, 2026, Kevin Flansburg and Michelle Chen published Cloudflare's account of a cache exposure on one Workers AI model, @cf/zai-org/glm-5.3. The window in the report is August 27 through September 9, 2026. Cloudflare says 155 responses were affected, and that every one of them answered a prompt longer than 681,280 tokens. It says some of those responses included text from another customer's request. It says other Workers AI models were not affected, that the model was restored on September 9, and that it found no evidence of deliberate exploitation. SGLang published a coordinated account the same day.

What this page will not describe

Cloudflare's post explains how a configuration and an engine bug interacted. This brief will not retell that path, and it will not describe a request that would reproduce it. The impact is the part a customer can use: one model, very long prompts, 155 responses, a closed window, a September 9 restoration, and an October 26 public report. Auto Router was a different Cloudflare document, dated in September, about a routing product. Origin Range was a cache rule for a different feature. Neither one is this incident.Auto Router's date.

The research paper is a different cache

AKV was filed on October 26 as a lab paper about a cache design and a savings table. A reader who sees KV in both headlines can merge them. They should not. AKV did not name a Cloudflare tenant, and Cloudflare's report does not cite that paper. The session brief is the privacy line that one person's request is not another person's context. That is the impact Cloudflare is describing, without this page repeating the mechanism. Asked directly, Is this the AKV paper? No. AKV is a research note about a cache design. This report is a production incident on one hosted model.The AKV paper.

Who was in scope

Cloudflare's scope is customers of that one GLM-5.3 endpoint whose prompts exceeded the length it names, during the August 27 to September 9 window. People who used a different Workers AI model are outside the report. People whose prompts were shorter than the threshold the report gives are outside it as well. The report does not publish customer names. Gadi Evron of Knostic is named as the reporter, on September 8, the day before the restoration. The document named in the sources is the record this page will keep.Origin Range.

SGLang's release

SGLang's October 26 post is the engine project's account of the same coordinated disclosure. The project also shipped v0.5.22, which the release page describes as bundling upstream fixes. A release tag is a version an operator of that engine can look up. It is not, by itself, a statement that every hosted copy of the model was safe on the day of the tag. Cloudflare's restoration date for the hosted model remains September 9. The advisory brief is the habit of keeping the vendor's date attached to the vendor's product.The advisory is the story.

Pingcap was a credential

Pingcap's filing was a vendor disabling a credential. Cloudflare's filing is a vendor describing crossed responses on one model. Both are distributed-cloud documents about something a customer did not intend to share. The fixes are different, and this page will not borrow one vendor's remedy as the other's. The short answer this page will stand on is this: No. The October 26 report covers @cf/zai-org/glm-5.3 only. 155 responses to prompts longer than 681,280 tokens were affected, between August 27 and September 9.Pingcap's credential.

The dates, kept in order

The exposure window ends September 9. The public report is October 26. A headline on October 27 can make the incident sound new. The customers in the 155 responses were in the August and September window. The news, on the day of the report, is that the account is now public and that the engine project published alongside it. SpaceDB's pin was a house datastore choice. It is not a cache incident. Asked directly, Who alerted Cloudflare? The report says Gadi Evron of Knostic reported the issue on September 8. Cloudflare says it found no evidence of deliberate exploitation.The SpaceDB pin.

What the company says it did not find

Cloudflare says it has no evidence the exposure was used deliberately. That is the company's finding, not a proof that every one of the 155 responses was harmless. The report is the place that finding lives. This page will not upgrade it to a clean bill, and it will not upgrade it to a breach list the company did not publish. Asked directly, How many responses does Cloudflare say were affected? 155, all of them answers to prompts longer than 681,280 tokens.

The record

The record is Cloudflare's October 26 post, SGLang's coordinated post, and the v0.5.22 release note. The model is @cf/zai-org/glm-5.3. The count is 155. The length threshold is 681,280 tokens. The window is August 27 to September 9. Other Workers AI models are out of scope, the company says. Asked directly, Which model does the Cloudflare report name? @cf/zai-org/glm-5.3. Cloudflare says other Workers AI models were not affected. Asked directly, What window does Cloudflare give? August 27 through September 9, 2026. The model was restored on September 9. The public report is dated October 26.

Sources

The reports this brief is filing. Futureweb did not republish them.

  1. Cloudflare
  2. SGLang
  3. SGLang 0.5.22

Questions

Which model does the Cloudflare report name?

@cf/zai-org/glm-5.3. Cloudflare says other Workers AI models were not affected.

How many responses does Cloudflare say were affected?

155, all of them answers to prompts longer than 681,280 tokens.

What window does Cloudflare give?

August 27 through September 9, 2026. The model was restored on September 9. The public report is dated October 26.

Who alerted Cloudflare?

The report says Gadi Evron of Knostic reported the issue on September 8. Cloudflare says it found no evidence of deliberate exploitation.

Is this the AKV paper?

No. AKV is a research note about a cache design. This report is a production incident on one hosted model.