A successful response can still contain the wrong customer’s data.

In its October 26 disclosure, Cloudflare describes a tenant-isolation failure in its GLM-5.3 Workers AI deployment. The company says 155 responses were affected between August 27 and September 9, all involving prompts longer than 681,280 tokens. It reports no evidence of deliberate exploitation and says no other Workers AI model was affected. This is a newly published account of an incident repaired in September.

The failure crossed a handoff between two sets of machines. One processes the prompt and builds a key-value cache, the model’s working record of that context. Another uses that cache to generate the answer. Cloudflare configured a 1,048,576-token context window, but the first set’s cache pool held only 681,280 tokens. Its long-context checks had accepted successful status codes without checking output coherence.

The SGLang maintainers’ coordinated account explains the second failure. A request marked for rejection could continue into the handoff. The answering machine could then read cache pages that the new request had never populated. A capacity problem became a data-boundary problem because rejection did not stop the work.

Cloudflare says it increased the pool and patched the rejection path before restoring the model on September 9. The maintainers also describe related fixes for cache pages reused while earlier transfers remain in flight. Their report distinguishes the incident’s admission failure from those broader lifetime problems.

For an operator, the useful test is concrete. Send distinct synthetic markers through separate test accounts, including requests near capacity limits, interrupted transfers and rejected work. Check that errors terminate cleanly and that markers never appear in the other account’s output. These are proposed verification steps, not tests performed by this publication.

Keep the deployment topology and engine version beside the result. A single-machine test cannot establish the behavior of a service that splits prompt processing from answer generation. The boundary must hold when a request fails, because that is when the next customer can inherit its unfinished state.