FAILURE MODES

Cloudflare: 155 AI Replies Held Other Customers' Data, All HTTP 200

Between Aug 27 and Sep 9, 155 replies from Cloudflare's hosted GLM-5.3 model on Workers AI contained data from other customers' requests. All 155 were replies to prompts longer than 681,280 tokens, and every one of them came back HTTP 200.

2026-10-27

Cloudflare says that between Aug 27 and Sep 9, 155 responses from its hosted GLM-5.3 model on Workers AI contained data from other customers’ requests. All 155 were replies to prompts longer than 681,280 tokens. Cloudflare says a small number of customers were affected, it found no evidence of deliberate exploitation, and no other Workers AI model was affected.

Every one of those 155 replies came back HTTP 200. The client got success, not an error, with a coherent reply.

The 40-second reel: 155 AI replies held other customers’ data.

What happened

The model was deployed with a 1,048,576-token context window, but the KV cache pool on its prefill GPUs, the GPUs that read the prompt, held only 681,280 tokens. Cloudflare knew about the mismatch before launch and expected the engine to return an error for prompts that did not fit.

In SGLang, the open-source engine it runs in split prefill/decode mode, an over-length prompt was marked rejected and swapped for a one-token placeholder, but the request went through anyway. Two sets of GPUs serve the model: one reads the prompt, one writes the reply. The reader rejected prompts too long for its memory, but sent the writer a one-token stand-in. The writer answered from memory never wiped of other users’ data.

Cloudflare’s own diagram shows the shape of it: the decode GPUs had reserved memory pages for the full prompt. Only one page was written, and the rest still held what earlier requests left there, other users’ prompts and partial generations. In the diagram’s example, decode reserved 8 pages, only P0 was written, and P1 to P7 still held earlier requests’ data.

The client got HTTP 200 instead of HTTP 400, with a coherent reply. Every request over the threshold behaved this way.

The testing gap

Cloudflare says its long-context tests before launch got success codes too. The tests checked status codes and speed, not the output. In Cloudflare’s own words, “we only validated HTTP response codes and engine performance at long contexts,” and “our long-context tests returned successful responses, so we did not believe this to be an issue.”

A 200 means the server reported success. It does not mean the answer is right.

The fix

Per Cloudflare, the prompt memory pool now holds about 2.04 million tokens, about twice the context window, and the engine now rejects over-length requests before the handoff. The SGLang patch shipped in v0.5.20, with follow-ups in v0.5.21 and v0.5.22. Cloudflare recommends v0.5.22 or later for anyone running disaggregated SGLang.

The cause, per Cloudflare, was its own configuration plus a bug in SGLang, not a fault in the model. The question the incident leaves for the rest of us is the one the reel ends on: do your tests check the answer, or just the 200?

What we did not verify

This is Cloudflare’s own account, published Oct 26, 2026. We did not test any of it. The 8-page diagram is an example from Cloudflare’s report, not a recording. The voiceover in the reel is AI-generated. Watch the reel on Threads or YouTube.

This post started as an Instagram post →

Discussion

Talk it through

Argue with us on Instagram.