SGLang Mooncake Backend Memory Exhaustion via ZMQ

Executive Summary

CVE-2026-112269 is a memory exhaustion vulnerability in SGLang through version 0.5.22. It allows network-adjacent attackers to send unauthenticated decode registration messages with unique session IDs and large nested-buffer frames to the Prefill rank ZMQ PULL socket. This exhausts Prefill memory until the scheduler is terminated due to out-of-memory conditions, causing a denial of service.

Useful 0 Not Useful 0

Detection Guidance

Monitor ZMQ PULL socket traffic on the Prefill rank for unauthenticated decode registration messages with unique mooncake_session_id values and large nested-buffer frames. Check for memory exhaustion events in the Prefill scheduler logs.

Useful 0 Not Useful 0

Impact Analysis

This vulnerability can cause system downtime by crashing the Prefill scheduler due to memory exhaustion. It may disrupt services relying on SGLang, such as AI inference workloads, until the system is restarted. Attackers can exploit it without authentication if they are on the same network as the vulnerable system.

Useful 0 Not Useful 0

Mitigation Strategies

Restrict network access to the Prefill rank ZMQ PULL socket, implement authentication for incoming messages, and set memory limits for the Prefill scheduler to prevent OOM conditions.

Useful 0 Not Useful 0