Blog

Cloudflare fixes GLM-5.3 cross-customer data exposure

Cloudflare’s GLM-5.3 incident connects shared-memory isolation, configuration checks and patch validation to practical Security+ learning.

· 4 min read

A service can return a successful response and still fail its most important security check: keeping one customer's information away from another. Cloudflare's October 26 disclosure makes that distinction concrete. Its GLM-5.3 model on Workers AI returned other customers' data in 155 responses between August 27 and September 9, 2026. Cloudflare says it fixed the issue before restoring the model on September 9, and found no evidence of deliberate exploitation.

The boundary was between requests

Cloudflare identified only the Workers AI model named @cf/zai-org/glm-5.3 as affected, not every model on its platform. All 155 affected responses followed prompts longer than 681,280 tokens, the pieces of text the model processes. That is a count of responses, not a count of compromised accounts. These scope limits come from Cloudflare's investigation; they should stay attached to any summary of the incident.

The model's advertised context window, meaning how much text a request could contain, was 1,048,576 tokens. But the temporary memory pool on the machines that processed incoming prompts could hold only 681,280. A request could therefore fit the advertised limit while exceeding the memory available for processing it. Cloudflare knew about the mismatch, but expected the software to reject requests it could not handle.

How shared memory crossed the customer boundary

The serving software, SGLang, used separate workers to process prompts and generate answers. It moved a request's KV cache, internal model state used to generate the response, between those workers. Cloudflare explains that this cache used a shared pool of memory pages. A later request was expected to overwrite the pages assigned to it.

For an over-length request, the prompt-processing worker marked the request for rejection and substituted a one-token placeholder. The handoff still proceeded. The answer-generating worker assumed that the full request's state had arrived, although much of the assigned memory still contained state from earlier requests. Its answer could consequently contain another customer's information. The SGLang maintainers' coordinated disclosure confirms this rejection-path failure and distinguishes it from related timing bugs involving interrupted transfers.

For a learner, the useful distinction is between permission to submit a request and isolation of the memory used to answer it. In this incident, a successful request was not evidence that its answer used only its own data. This was a serving-time memory-handling failure, not evidence that training had taught the model those private conversations.

Why a green availability check was not enough

Cloudflare's pre-launch long-context tests checked successful web response codes and performance, but did not validate the output's correctness and coherence. The report identifies that testing gap directly. A successful web response showed that the service answered; it did not prove that the customer boundary held.

Cloudflare corrected both layers. It enlarged the memory pool so valid requests fit, and patched the rejection path so rejected requests stopped before the handoff. Before restoring access, it tested the changes against the researcher's reproduction and its own canary-marker tests, which use recognizable markers to detect unintended data appearing in a response. It also added deployment checks that compare the running engine's pool capacity with the configured context length.

That is a useful way to distinguish configuration enforcement from patching. Matching the capacity removes this trigger; fixing the rejection path addresses unsafe behavior when a request is rejected. Neither description should be reduced to “add more memory.” Cloudflare's operator guidance recommends SGLang 0.5.22 or later and further configuration and transfer checks. The upstream disclosure lists additional backend-specific follow-ups, so that minimum is not a blanket assurance about every deployment or every cache bug.

Turn the incident into a Security+ exercise

CompTIA's SY0-701 objectives summary includes confidentiality, isolation, configuration enforcement, patching, vulnerability validation and incident response. Those are the connections to study here, rather than memorizing one model's token limit. This is a learning example, not a prediction that this incident will appear on an exam.

Write three short answers using the disclosure as your evidence. First, identify the failed security property: another customer's data appeared in a response, so confidentiality was breached. Second, separate the two corrective controls: capacity enforcement and the software fix to rejection handling. Third, describe validation that checks the customer boundary rather than just whether the service responds. For an authorized lab exercise, use invented customer data and distinct markers, never real private records.

Then compare those answers with the concepts in our Security+ SY0-701 study guide. The point is to practice explaining why a control fits the failure. A running service and a correctly isolated service are different claims, and this incident shows why both need evidence.

Read the SY0-701 study guide

Sources