Cloudflare said that a GLM 5.3 configuration mismatch in Workers AI led to an SGLang error. As a result, KV-cache pages ended up open between different requests.
The company also stated that it remedied the issue with fixes in SGLang. The available description does not specify the incident period, the number of affected requests, or the nature of the data that could have been exposed.
The significance of the incident lies in the risk of cross-request data access in systems where KV-cache is used to speed up model operation. However, the source is presented only as a brief Cloudflare synopsis without independent confirmation and details of the technical analysis.
editorial commentary
Why it matters
Probable consequence — increased attention to cache isolation and configuration compatibility in model-inference systems. The next observable signal will be Cloudflare publishing technical details, scale of affected requests, and a description of fixes. Substantial uncertainty remains due to the absence of full text and independent verification.