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-2026-112082 Overview
CVE-2026-112082 is a denial of service vulnerability in sgl-project SGLang versions up to 0.5.19. The flaw resides in the Scheduler.expert_distribution_handle function within python/sglang/srt/eplb/expert_distribution.py, part of the Expert-Distribution Recording Endpoint component. Remote attackers can trigger the issue over the network without authentication or user interaction, causing service disruption. A public exploit is referenced in the advisory, and the fix remains pending acceptance in an upstream pull request at the time of publication.
Critical Impact
Unauthenticated remote attackers can disrupt SGLang inference service availability by abusing the Expert-Distribution Recording Endpoint, impacting hosted large language model (LLM) workloads.
Affected Products
- sgl-project SGLang up to version 0.5.19
- The python/sglang/srt/eplb/expert_distribution.py module
- Deployments exposing the Expert-Distribution Recording Endpoint
Discovery Timeline
- 2026-10-19 - CVE-2026-112082 published to the National Vulnerability Database (NVD)
- 2026-10-19 - Last updated in NVD database
Technical Details for CVE-2026-112082
Vulnerability Analysis
SGLang is a serving framework for large language models and vision-language models. The Expert-Distribution Recording Endpoint exposes operational hooks used to inspect and record expert routing behavior in mixture-of-experts (MoE) workloads. The Scheduler.expert_distribution_handle function processes requests to this endpoint without adequate controls over resource usage or state transitions.
The weakness is classified as improper resource shutdown or release [CWE-404]. An attacker reaching the endpoint can manipulate handler invocations in a way that leaves the scheduler unable to continue servicing legitimate inference requests. Because SGLang often hosts production LLM inference endpoints, a sustained attack results in degraded throughput or a full outage of the model serving process.
Root Cause
The handler does not enforce proper resource lifecycle management when the Expert-Distribution Recording Endpoint is invoked. Repeated or malformed interaction with expert_distribution_handle leaves scheduler resources in an unrecoverable state, exhausting capacity to process new requests. The upstream project is tracking a fix in GitHub Pull Request #42098, which has not yet been merged.
Attack Vector
Exploitation requires only network access to the SGLang service. No credentials and no user interaction are needed. An attacker sends crafted requests to the Expert-Distribution Recording Endpoint, triggering the faulty handler path. Public exploit details are referenced through VulDB entry 421761, increasing the likelihood of opportunistic abuse against exposed deployments.
No verified exploit code is published in this advisory. For technical reproduction details, see the upstream GitHub Issue #42071 and the proposed fix in GitHub Pull Request #42098.
Detection Methods for CVE-2026-112082
Indicators of Compromise
- Repeated HTTP requests to the Expert-Distribution Recording Endpoint from untrusted sources
- Abrupt termination or hang of SGLang scheduler worker processes without a corresponding administrative action
- Sudden drops in inference throughput and rising request latency on SGLang-fronted services
- Scheduler log entries referencing expert_distribution_handle immediately before service degradation
Detection Strategies
- Instrument SGLang with request-level logging on the Expert-Distribution Recording Endpoint and alert on anomalous call rates
- Correlate scheduler process restarts with inbound network activity targeting SGLang management paths
- Baseline normal client sources for the endpoint and alert on requests from unexpected IP ranges
Monitoring Recommendations
- Track process health and memory of SGLang scheduler workers and alert on repeated crashes
- Monitor north-south traffic to inference hosts for scanning activity against SGLang HTTP paths
- Review reverse-proxy and ingress logs for calls to expert-distribution routes from outside expected administrative networks
How to Mitigate CVE-2026-112082
Immediate Actions Required
- Restrict network exposure of SGLang services so the Expert-Distribution Recording Endpoint is unreachable from untrusted networks
- Place SGLang behind an authenticated reverse proxy or service mesh with per-route access control
- Disable the Expert-Distribution Recording Endpoint in production deployments where it is not operationally required
- Monitor the upstream GitHub SGLang Repository for the merge of Pull Request #42098 and apply the patch once released
Patch Information
At the time of publication, no fixed SGLang release is available. The proposed fix is tracked in GitHub Pull Request #42098 and awaits acceptance. Operators should subscribe to the pull request and the associated GitHub Issue #42071 for status updates, and upgrade immediately once a release containing the fix is published. Additional advisory details are available on VulDB CVE-2026-112082.
Workarounds
- Enforce IP allow-listing at the network or ingress layer to limit access to SGLang administrative endpoints
- Apply rate limiting on requests reaching the Expert-Distribution Recording Endpoint to blunt resource exhaustion attempts
- Run SGLang under a process supervisor configured for rapid automatic restart to shorten outage windows while a patch is pending
- Segment inference infrastructure from general corporate and internet-facing networks using firewall policy
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

