SGLang versions 0 through v0.5.15 are susceptible to an unauthenticated model weight exfiltration vulnerability that poses a high risk to intellectual property.

This vulnerability involves a missing authentication check (CWE-306) in SGLang, where the application exposes two endpoints that facilitate unauthorized distributed weight broadcasting via NCCL. The vulnerability is fully unauthenticated, meaning any remote attacker can trigger the data transfer mechanism to exfiltrate sensitive model weights.

The exploitation of this vulnerability results in the unauthorized exfiltration of proprietary model weights, which constitutes a significant loss of intellectual property. Given the CVSS score of 7.5, this high-severity flaw enables attackers to steal core AI assets without requiring any prior access or credentials, potentially leading to severe competitive disadvantage and reputational harm.

Immediate Action: Review the official SGLang security advisory for the latest version and apply the update immediately. If a patch is not yet available, ensure API keys are configured and restrict network access to the affected endpoints.

Proactive Monitoring: Monitor network traffic and server logs for unusual requests to endpoints related to distributed weight broadcasting or NCCL operations.

Compensating Controls: Deploy a Web Application Firewall or network access control list to restrict external access to the SGLang management endpoints.

Public Exploit Available: No.

Organizations utilizing SGLang must treat this vulnerability with high urgency due to the potential for complete loss of model confidentiality. Administrators should verify their current deployment version and restrict access to the affected endpoints while awaiting vendor-provided patches. Prioritizing the implementation of strong authentication and network segmentation will provide necessary defense-in-depth against this threat.