Cloudflare reports cross-request KV cache exposure in GLM-5.3 on Workers AI and describes remediation with SGLang
Cloudflare published a post describing a configuration mismatch in the GLM-5.3 model on Workers AI that triggered an SGLang bug, which exposed KV cache pages across requests. The post also describes the steps Cloudflare took to address the issue.
Why it matters
Builders using hosted model inference on shared infrastructure should know that KV cache data can leak between requests when a serving-stack bug or configuration mismatch occurs. This affects how much trust to place in multi-tenant inference for sensitive prompts and outputs.
Who should care
Developers building on Workers AI, teams sending user or customer data to hosted LLM endpoints, security and compliance leads, and teams self-hosting SGLang.
What you can do
Read the full Cloudflare post, check whether your application sends sensitive data to GLM-5.3 on Workers AI, and ask Cloudflare to confirm which remediation applies to your deployment.