Skip to main content
Vulnerability Database/CVE-2026-112269

CVE-2026-112269: SGLang Mooncake Backend DoS Vulnerability

CVE-2026-112269 is a denial of service vulnerability in SGLang through version 0.5.22 with Mooncake disaggregation backend that allows network-adjacent attackers to exhaust memory. This post covers technical details, affected versions, impact assessment, and mitigation strategies.

Published:

Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.

Sources: National Vulnerability Database, CISA KEV, Exploit-DB

CVE-2026-112269 Overview

CVE-2026-112269 is a denial-of-service vulnerability in SGLang through version 0.5.22 when deployed with the Mooncake disaggregation backend. The Prefill rank ZMQ PULL socket accepts decode registration messages without authentication. Network-adjacent attackers can flood the socket with unique mooncake_session_id registrations and large nested-buffer frames. The resulting unbounded allocations exhaust Prefill process memory until the Linux Out-Of-Memory (OOM) killer terminates the scheduler. The flaw is categorized under CWE-770 (Allocation of Resources Without Limits or Throttling).

Critical Impact

An unauthenticated adjacent-network attacker can crash the SGLang Prefill scheduler, disrupting LLM inference service availability for all downstream clients.

Affected Products

  • SGLang versions through 0.5.22
  • Deployments using the Mooncake disaggregation backend
  • Prefill rank nodes exposing the ZMQ PULL socket on reachable network interfaces

Discovery Timeline

  • 2026-10-19 - CVE-2026-112269 published to NVD
  • 2026-10-19 - Last updated in NVD database

Technical Details for CVE-2026-112269

Vulnerability Analysis

SGLang's Mooncake disaggregation backend splits large language model inference between Prefill and Decode ranks. Decode workers register with Prefill nodes over a ZeroMQ PULL socket to coordinate key-value cache transfers. The registration handler in python/sglang/srt/disaggregation/mooncake/conn.py deserializes incoming frames and stores per-session state keyed by mooncake_session_id.

The handler performs no authentication, no sender validation, and no cap on the number of concurrent registrations. It also accepts nested-buffer frames without bounding their size. An attacker on an adjacent network segment can repeatedly send crafted registration messages, each with a unique session identifier, causing the session table to grow without limit. Large payload frames amplify memory consumption per message.

Root Cause

The root cause is missing resource throttling on an unauthenticated ingress path [CWE-770]. The Prefill rank trusts any peer reaching its ZMQ socket as a legitimate Decode worker. There is no quota on mooncake_session_id entries, no size limit on buffer frames, and no sender identity binding. See the GitHub registration handler code and buffer handling code for the vulnerable logic.

Attack Vector

The attack vector is adjacent network. An attacker with access to the subnet hosting the Prefill rank sends a sustained stream of fake decode registrations to the exposed ZMQ PULL endpoint. Each message carries a unique mooncake_session_id and oversized nested buffers. The Prefill process allocates memory for each entry until the kernel OOM-killer terminates the scheduler, halting all inference workloads. Additional technical context is available in the VulnCheck advisory and GitHub Issue #45012.

Detection Methods for CVE-2026-112269

Indicators of Compromise

  • Rapid, sustained inbound ZMQ traffic to Prefill rank ports from unexpected hosts on the inference subnet
  • Prefill process resident set size (RSS) climbing linearly without a corresponding increase in client inference requests
  • Kernel oom-killer log entries terminating the SGLang scheduler process
  • Large volumes of distinct mooncake_session_id values in SGLang logs within short time windows

Detection Strategies

  • Alert on anomalous session registration rates per source IP against the Prefill ZMQ endpoint
  • Correlate scheduler restart events with preceding spikes in Prefill memory utilization
  • Baseline normal Decode-to-Prefill peer counts and flag deviations above a threshold

Monitoring Recommendations

  • Export process-level memory metrics for SGLang Prefill workers to a time-series backend
  • Capture kernel dmesg output for OOM events and forward to centralized logging
  • Monitor firewall and flow logs for ZMQ port traffic originating outside the authorized Decode rank set

How to Mitigate CVE-2026-112269

Immediate Actions Required

  • Restrict network reachability of the Prefill ZMQ PULL socket to known Decode rank IP addresses using host firewalls or network policies
  • Place SGLang disaggregation traffic on an isolated management VLAN or private overlay network
  • Set operating-system memory limits (cgroups or systemd MemoryMax) on the Prefill scheduler to force graceful failure instead of host-wide impact
  • Monitor the SGLang project for a patched release addressing the Mooncake registration handler

Patch Information

At the time of publication, no fixed version is identified in the NVD record. The affected code paths referenced in the advisory are present through SGLang 0.5.22. Track the SGLang GitHub repository and GitHub Issue #45012 for upstream fixes that add authentication, session caps, and buffer-size validation on the Prefill ingress path.

Workarounds

  • Deploy a reverse proxy or mutual TLS tunnel in front of the ZMQ socket to enforce peer authentication
  • Apply iptables or nftables rate limits on the Prefill rank's ZMQ port to cap registration throughput per source
  • Run Prefill workers under container runtime memory limits so that OOM termination is scoped and auto-restarted without impacting the host
  • Disable the Mooncake disaggregation backend where disaggregated inference is not required
bash
# Configuration example: restrict ZMQ PULL port to authorized Decode ranks
# Replace 10.0.0.0/24 with the Decode subnet and 5557 with the Prefill ZMQ port
iptables -A INPUT -p tcp --dport 5557 -s 10.0.0.0/24 -j ACCEPT
iptables -A INPUT -p tcp --dport 5557 -j DROP

# Enforce a memory ceiling on the Prefill scheduler via systemd
sudo systemctl set-property sglang-prefill.service MemoryMax=48G

Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.

Sources: National Vulnerability Database, CISA KEV, Exploit-DB

Experience the Most Advanced Cybersecurity Platform

See how the world’s most intelligent, autonomous cybersecurity platform can protect your organization today and into the future.