CVE-2026-15976

Analyzed Analyzed - Analysis Complete

Remote Code Execution in SGLang via Malicious Model Weights

Publication date: 2026-07-30

Last updated on: 2026-08-04

Assigner: CERT/CC

Description

SGLang contains a RCE vulnerability when attempting to load model weights from a HuggingFace repository, specifically within the /update_weights_from_disk, where torch.load(..., weights_only=False) fallback enables pickle deserialization of .bin files.

CVSS Scores

EPSS Scores

Probability:
Percentile:

Meta Information

Published

2026-07-30

Last Modified

2026-08-04

Generated

2026-08-20

AI Q&A

2026-07-31

EPSS Evaluated

2026-08-18

NVD

CVE-2026-15976

EUVD

EUVD-2026-51264

Affected Vendors & Products

Vendor	Product	Version / Range
lmsys	sglang	to 0.5.15 (inc)

Helpful Resources

Exploitability

CWE

KEV

CWE ID	Description
CWE-502	The product deserializes untrusted data without sufficiently ensuring that the resulting data will be valid.

Attack-Flow Graph

Executive Summary

SGLang has a remote code execution vulnerability when loading model weights from HuggingFace. The issue occurs in the /update_weights_from_disk function where torch.load uses weights_only=False, allowing pickle deserialization of .bin files which can execute arbitrary code.

Detection Guidance

This vulnerability involves unsafe deserialization of .bin files via torch.load with weights_only=False in SGLang. To detect it, inspect Python processes loading HuggingFace models for suspicious torch.load calls with weights_only=False. Check for unexpected .bin files in model directories or network traffic to untrusted HuggingFace repositories.

Impact Analysis

This vulnerability could allow attackers to execute arbitrary code on your system if you load a malicious model from HuggingFace. This could lead to data theft, system compromise, or further network infiltration.

Compliance Impact

This vulnerability could lead to unauthorized access to personal or sensitive data, violating GDPR and HIPAA requirements for data protection and confidentiality. Organizations may face fines or legal consequences if exploited.

Mitigation Strategies

Avoid loading model weights from untrusted HuggingFace repositories. Use only verified sources and disable pickle deserialization by setting weights_only=True in torch.load calls.

Hi! I’m here to help you understand CVE-2026-15976. Ask me anything about the vulnerability, its impact, or mitigation strategies.

0/70