Remote DoS in SGLang Scheduler Affects Versions Up to 0.5.19

A denial-of-service vulnerability has been identified in the sgl-project SGLang software, specifically impacting versions up to 0.5.19. The flaw resides in the Scheduler.expert_distribution_handle function within the Expert-Distribution Recording Endpoint component. According to the National Vulnerability Database record, this issue allows an attacker to launch a remote attack that results in service disruption. The exploit is publicly available and might be used, increasing the potential for immediate abuse against exposed instances.

The risk

The vulnerability is classified as HIGH severity with a CVSS score of 8.7. It maps to CWE-404, which relates to improper resource control leading to resource exhaustion or denial of service. The specific function affected, located in the file python/sglang/srt/eplb/expert_distribution.py, handles expert distribution logic. Manipulation of this endpoint enables remote attackers to trigger the denial-of-service condition without requiring local access or authentication. The public availability of the exploit means that attackers can leverage this flaw against exposed instances.

Who is exposed

Organizations running sgl-project SGLang versions 0.5.0, 0.5.1, 0.5.2, or 0.5.3 are directly exposed to this risk. The National Vulnerability Database record explicitly lists these versions as affected, along with any releases up to 0.5.19. Teams managing infrastructure using this specific software stack should verify their current version numbers. Any environment where the Expert-Distribution Recording Endpoint is accessible from untrusted networks faces heightened risk due to the remote nature of the attack vector.

Cost and compliance

The source material does not provide specific financial impact data or compliance implications for this vulnerability. The National Vulnerability Database record focuses strictly on the technical details of the flaw, its severity rating, and the affected software versions. There is no mention of regulatory fines, data breach costs, or specific compliance frameworks triggered by this denial-of-service vector. Organizations must assess internal risk tolerance and service-level agreement impacts independently, as the public record limits its scope to the technical existence of the vulnerability and the status of the remediation effort.

What to watch

Security teams should monitor for unusual traffic patterns targeting the expert distribution endpoints of SGLang instances. Watch for sudden drops in service availability or increased error rates in the scheduler components. Since the exploit is public, automated attacks may begin appearing in logs. Teams should also track the status of the pending pull request, as the fix has not yet been officially released or merged into the main codebase.

Signals worth monitoring

Teams usually see a spike in connection attempts to specific API endpoints when this type of denial-of-service attack begins. Check whether the scheduler component is receiving malformed requests that trigger abnormal resource consumption. Look for logs indicating repeated failures in the expert distribution handling routines, which may precede total service unavailability. Verify that network monitoring tools are capturing traffic directed at the recording endpoint, as this is the primary vector. Ensure that alerting thresholds for CPU and memory usage in the scheduler process are sensitive enough to catch rapid exhaustion before it impacts broader service availability.

What to do and how to stay safe: SGLang

• Inventory all SGLang instances to confirm if they are running versions 0.5.0 through 0.5.19.
• Restrict network access to the Expert-Distribution Recording Endpoint to trusted internal IPs only.
• Monitor system resources for sudden spikes that indicate ongoing exploitation attempts.
• Prepare to apply the patch once the vendor accepts the pending pull request and releases an update.

Step-by-step guide: Vulnerability Prioritization: Fix What Actually Matters First

General security guidance from the SecOps Signal newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

Which specific versions of SGLang are vulnerable to CVE-2026-112082?

Versions 0.5.0, 0.5.1, 0.5.2, 0.5.3, and all versions up to 0.5.19 are affected.

Is there a confirmed patch available for this vulnerability yet?

No, a pull request to fix the issue awaits acceptance, so no official patch is confirmed.

What is the severity rating of this vulnerability?

The CVSS score is 8.7, which is rated as HIGH severity.