Skip to main content
Vulnerability Database/CVE-2026-112270

CVE-2026-112270: SGLang Mooncake Backend DoS Vulnerability

CVE-2026-112270 is a denial of service vulnerability in SGLang through version 0.5.22 with Mooncake disaggregation backend that allows unauthenticated attackers to exhaust memory. This post explains its impact, affected versions, and mitigation steps.

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-2026-112270 Overview

CVE-2026-112270 is an uncontrolled resource allocation vulnerability [CWE-770] in SGLang through version 0.5.22 when deployed with the Mooncake disaggregation backend. The flaw resides in the handle_staging_rsp() function within the staging handler. Unauthenticated network-adjacent attackers can exhaust Prefill memory and trigger an out-of-memory (OOM) kill of the scheduler. Attackers retrieve the rank address from the /route endpoint, then send crafted STAGING_RSP messages containing oversized chunk_idx values under fresh room IDs. The vulnerability enables denial-of-service against large language model (LLM) inference infrastructure built on SGLang.

Critical Impact

Unauthenticated adjacent attackers can OOM-kill the SGLang scheduler, disrupting LLM inference services.

Affected Products

  • SGLang through version 0.5.22
  • SGLang deployments configured with the Mooncake disaggregation backend
  • Prefill nodes exposing the /route endpoint on reachable networks

Discovery Timeline

  • 2026-10-19 - CVE-2026-112270 published to the National Vulnerability Database (NVD)
  • 2026-10-19 - Last updated in NVD database

Technical Details for CVE-2026-112270

Vulnerability Analysis

SGLang supports a disaggregated serving mode where Prefill and Decode workers coordinate over the Mooncake transfer engine. The handle_staging_rsp() function processes inbound STAGING_RSP messages that reference staging buffers by room ID and chunk index. The handler allocates or indexes memory structures based on the attacker-supplied chunk_idx field without validating that the value falls within an expected range. An attacker who submits very large chunk_idx values forces unbounded memory allocation on the Prefill node. Repeated requests under fresh room IDs cause continuous allocation until the Linux OOM killer terminates the scheduler process. The result is a reliable denial-of-service against inference workloads.

Root Cause

The root cause is missing bounds validation on untrusted input to a memory-sizing operation, classified as [CWE-770] Allocation of Resources Without Limits or Throttling. The handler trusts the chunk_idx field in STAGING_RSP messages rather than enforcing per-room or per-connection ceilings.

Attack Vector

Exploitation requires network adjacency to the Prefill node and no authentication. The attacker first queries the /route endpoint to discover the Mooncake rank address. The attacker then sends STAGING_RSP messages with inflated chunk_idx values, assigning each a new room ID to bypass any per-room reuse logic. See the GitHub SGLang staging handler code and Mooncake connection code for the vulnerable logic.

// No verified exploit code is published. See the VulnCheck advisory
// and the GitHub source links for technical details.

Detection Methods for CVE-2026-112270

Indicators of Compromise

  • Scheduler processes terminated by the Linux OOM killer with corresponding dmesg entries referencing the SGLang Python process.
  • Unexpected spikes in resident set size on Prefill nodes without matching inference workload increases.
  • Inbound traffic to the Mooncake rank address originating from clients that did not first complete a legitimate inference workflow.
  • Access patterns against the /route endpoint from unknown source addresses on the adjacent network.

Detection Strategies

  • Instrument the handle_staging_rsp() call path to log chunk_idx values exceeding configured maximums.
  • Monitor Prefill node memory consumption trends and alert on rapid growth tied to new room IDs.
  • Correlate /route endpoint queries with subsequent STAGING_RSP traffic from the same source.

Monitoring Recommendations

  • Export process-level memory metrics for the SGLang scheduler to a time-series platform and alert on sustained increases.
  • Collect kernel OOM-killer events and forward them to a centralized log store for correlation with network telemetry.
  • Track the rate of new room ID creation per source address and alert on anomalous bursts.

How to Mitigate CVE-2026-112270

Immediate Actions Required

  • Restrict network access to the Mooncake rank address and the /route endpoint using firewall rules or service-mesh policies.
  • Place Prefill nodes behind an authenticating reverse proxy or mTLS gateway accessible only to Decode peers.
  • Set memory limits (cgroup or Kubernetes resource caps) on scheduler pods to contain allocation attempts and ease recovery.
  • Monitor the SGLang issue tracker for an official fix and plan upgrade windows.

Patch Information

At the time of publication, no fixed version is listed beyond SGLang 0.5.22. Track the VulnCheck advisory and GitHub SGLang Issue #45014 for release updates and apply upstream patches as they become available.

Workarounds

  • Deploy SGLang Prefill and Decode components on an isolated management network that excludes untrusted tenants.
  • Front the /route endpoint with an authenticating proxy and strip STAGING_RSP messages from unapproved sources.
  • Add input validation at a sidecar layer that rejects STAGING_RSP messages carrying chunk_idx values above a sane maximum.
  • Enforce container memory limits so a single OOM event restarts the scheduler quickly without impacting the host.
bash
# Example: restrict Mooncake rank port to Decode peers only
iptables -A INPUT -p tcp --dport <mooncake_rank_port> -s <decode_peer_cidr> -j ACCEPT
iptables -A INPUT -p tcp --dport <mooncake_rank_port> -j DROP

# Example: enforce a memory ceiling on the SGLang scheduler container
docker run --memory=32g --memory-swap=32g sglang:0.5.22

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.