What Happened
Cloudflare reported a security issue affecting the GLM 5.3 model running on its Workers AI infrastructure. A configuration mismatch in the GLM 5.3 deployment triggered a bug within the SGLang inference framework, leading to cross-request Key-Value (KV) cache exposure.
Key Details
- Affected Service: GLM 5.3 model deployed on Cloudflare Workers AI.
- Technical Cause: A configuration mismatch activated a bug in the SGLang serving framework, allowing KV cache pages to leak across isolated request contexts.
- Status: Cloudflare has implemented remediation measures in coordination with SGLang.
Note: The source context does not specify the exact window of vulnerability, the volume of affected requests, or whether sensitive prompt data was exposed to third parties.
Why It Matters
In Large Language Model (LLM) architecture, KV caching optimizes processing speed by storing intermediate attention states during token generation. If KV cache memory pages are improperly isolated between requests, prompt context or generated responses from one user could potentially spill over into another user's session. Maintaining strict isolation at the inference server layer is crucial for multi-tenant AI platforms.
