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-2026-112083 Overview
CVE-2026-112083 is a denial of service vulnerability affecting the sgl-project SGLang inference framework through version 0.5.19. The flaw resides in the WeightExporter.send_weights_to_remote_instance function within python/sglang/srt/model_executor/model_runner_components/weight_exporter.py. An attacker can manipulate the group_name argument to trigger a denial of service condition in the Weight Exporter component. The vulnerability is remotely exploitable and does not require authentication or user interaction. Public exploit code has been released, increasing the risk to exposed SGLang deployments. The upstream pull request addressing the issue is pending acceptance at the time of disclosure.
Critical Impact
Remote unauthenticated attackers can trigger a denial of service in SGLang Weight Exporter by manipulating the group_name argument, disrupting model-serving availability.
Affected Products
- sgl-project SGLang versions up to and including 0.5.19
- Weight Exporter component (weight_exporter.py)
- Deployments exposing the send_weights_to_remote_instance interface to untrusted networks
Discovery Timeline
- 2026-10-19 - CVE-2026-112083 published to NVD
- 2026-10-19 - Last updated in NVD database
Technical Details for CVE-2026-112083
Vulnerability Analysis
SGLang is a serving framework for large language models that supports distributed weight synchronization across instances. The Weight Exporter component coordinates weight transfer between model-serving processes. The send_weights_to_remote_instance function accepts a group_name parameter used to identify communication groups.
Improper handling of the group_name argument allows a remote attacker to disrupt normal operation of the Weight Exporter. The weakness is classified under [CWE-404] (Improper Resource Shutdown or Release), indicating resource mismanagement that leads to loss of service availability. Successful exploitation prevents legitimate weight-synchronization operations and may require operator intervention to restore service.
Root Cause
The root cause is insufficient validation and resource management around the group_name input processed by WeightExporter.send_weights_to_remote_instance. Supplying crafted values causes the component to enter a state where resources are not properly released, consistent with [CWE-404]. Review the GitHub Issue #42073 and Pull Request #44314 for upstream analysis.
Attack Vector
The attack vector is network-based with low complexity, no privileges required, and no user interaction. An attacker reaches the exposed Weight Exporter endpoint over the network and submits a crafted group_name value. The exploit has been publicly released, so proof-of-concept code is available to any threat actor targeting SGLang deployments.
See the vulnerability description in the VulDB CVE-2026-112083 record for additional technical context. No verified exploit code is included here; refer to upstream references for implementation specifics.
Detection Methods for CVE-2026-112083
Indicators of Compromise
- Unexpected termination, hangs, or restarts of SGLang model-serving workers tied to weight-synchronization operations
- Repeated inbound requests to the Weight Exporter endpoint containing anomalous or malformed group_name values
- Errors in SGLang logs referencing send_weights_to_remote_instance or failed group registration
Detection Strategies
- Inventory running SGLang instances and confirm the installed version against the vulnerable range (≤ 0.5.19)
- Inspect application logs for WeightExporter errors and correlate with network sources making the requests
- Deploy network telemetry to flag unauthenticated connections to Weight Exporter ports from outside trusted ranges
Monitoring Recommendations
- Alert on SGLang process crashes, restarts, or resource-exhaustion events on inference hosts
- Monitor availability of model-serving endpoints with external synthetic checks to detect degraded service
- Centralize SGLang logs in a SIEM and build detections for repeated failures tied to group_name parameter handling
How to Mitigate CVE-2026-112083
Immediate Actions Required
- Restrict network access to SGLang Weight Exporter interfaces so only trusted internal hosts can reach them
- Place SGLang services behind an authenticated reverse proxy or service mesh enforcing mutual TLS
- Rate-limit requests to weight-transfer endpoints to reduce exposure to repeated malicious input
- Monitor the SGLang GitHub project for the merged fix and release a patched version
Patch Information
At publication, the upstream fix tracked in Pull Request #44314 was pending acceptance. Operators should track the pull request and apply the patch once it is merged and included in an official SGLang release. Until a fixed release is available, apply the compensating network and access controls described below.
Workarounds
- Isolate SGLang nodes on a dedicated management network segment unreachable from the public internet
- Enforce firewall rules allowing Weight Exporter traffic only from known model-serving peers
- Disable or disallow external invocation of send_weights_to_remote_instance where distributed weight transfer is not required
- Validate and allowlist expected group_name values at an upstream proxy if feasible
# Example: restrict Weight Exporter access with iptables to a trusted subnet
iptables -A INPUT -p tcp --dport <sglang_port> -s 10.0.0.0/24 -j ACCEPT
iptables -A INPUT -p tcp --dport <sglang_port> -j DROP
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

