Loading
Loading articles
By Kevin Flansburg

AI summary
Cloudflare identified a tenant isolation vulnerability in the GLM-5.3 Workers AI model caused by a mismatch between the allocated KV cache pool and the model's context window. The issue occurred when the SGLang inference engine improperly handled over-length requests in disaggregated prefill/decode mode, leading to cross-request KV cache exposure. The incident was remediated by patching the inference engine and Cloudflare is now applying formal verification methods to ensure the safety of its distributed KV cache transfer protocols.
Key points