A tenant isolation vulnerability in the GLM-5.3 Workers AI model allowed cross-request KV cache exposure for prompts exceeding 681,280 tokens. The issue, caused by an SGLang inference engine bug, resulted in 155 responses containing data from previous users before being remediated on September 9.
HOW THIS AFFECTS YOU
●
builderYou must verify tenant isolation when using high-context models on shared inference engines.
●
policyThis highlights the critical need for rigorous security auditing of multi-tenant inference infrastructure.