Cross-request KV cache exposure in the GLM-5.3 Workers AI model (155 responses) (blog.cloudflare.com)

A significant security issue was discovered in the GLM-5.3 Workers AI model, where requests with lengthy prompts inadvertently returned responses containing data from other users' previous sessions. This cross-request key value (KV) cache exposure, reported by security researcher Gadi Evron, affected 155 responses between August 27 and September 9, 2026, before remediation was enacted. Cloudflare promptly identified the root cause as a misconfiguration of the open-source inference engine SGLang and a bug within it that mishandled prompts exceeding the KV cache size. The incident highlights critical vulnerabilities in tenant isolation for AI models and underscores the importance of rigorous testing in configurations, particularly for high-performance large language models. Cloudflare's immediate response included aligning the KV cache size with the model's context window, effectively eliminating the exposure, and implementing a patch that ensures proper error handling for rejected requests. These developments not only fix the existing flaw but also reinforce the security posture across all models utilizing SGLang, prompting other operators to update their systems to prevent similar issues.