Cloudflare disclosed that the GLM-5.3 model on its Workers AI service returned fragments of other customers’ requests in 155 responses. The problem occurred between 27 August and 9 September 2026 and involved very long prompts. The cause was a KV cache pool smaller than the model’s context window, combined with a bug in the SGLang engine that let requests which should have been rejected read stale data. The company fixed the configuration and the engine, and SGLang version 0.5.22 includes related hardening.

What it means: This is one of the first publicly documented cases where the serving infrastructure of a language model, not the user or the model itself, caused data to leak between customers. The engines that run models keep intermediate results in memory for speed and share that memory across requests, so a mistake at the boundaries can expose someone else’s data. The number of affected responses was small and Cloudflare gave a detailed explanation, which is to its credit. For Greek businesses that send customer documents, invoices or personal data to AI services, the lesson is that risk is not only about what the model does but also about how the provider runs it. It is worth checking whether the provider commits to data isolation, how it reports incidents and whether dedicated capacity is available for sensitive workloads. Under GDPR, an exposure like this may be a breach that has to be assessed, so processing agreements should require prompt notice. Removing or pseudonymising names, tax numbers and account numbers before sending text to external models is also good practice.

Read the original on Cloudflare Blog

Related service: Custom software

Related stories