Critical RCE in SGLang via Malicious GGUF Models
A critical remote code execution (RCE) vulnerability, tracked as CVE-2026-5760 with a CVSS score of 9.8, has been disclosed in SGLang. The Hacker News reports this as a command injection flaw, allowing arbitrary code execution through specially crafted GGUF model files.
SGLang, an open-source serving framework, is designed for high-performance large language model (LLM) serving. This vulnerability means that any system processing untrusted GGUF models via SGLang is at severe risk. Attackers could embed malicious commands within these model files, which would then execute on the host system when the model is loaded or processed.
For defenders, this is a significant supply chain and execution risk. The attackerβs calculus is straightforward: leverage the trust placed in AI models to achieve deep system compromise. CISOs must understand that AI model files are not benign data objects; they are executable code in a different format, and this vulnerability proves it. Treat any external AI model with extreme suspicion.
What This Means For You
- If your organization uses SGLang to serve or process GGUF models, you are directly exposed. Immediately identify all instances of SGLang deployment and assess your model sourcing. You must validate the integrity and origin of every GGUF model file before it touches your SGLang environment. Patching or implementing robust input validation for GGUF models is not optional.
Related ATT&CK Techniques
π‘οΈ Detection Rules
3 rules Β· 6 SIEM formats3 detection rules auto-generated for this incident, mapped to MITRE ATT&CK. Sigma YAML is free β export to any SIEM format via the Intel Bot.
SGLang GGUF Command Injection RCE
Indicators of Compromise
| ID | Type | Indicator |
|---|---|---|
| CVE-2026-5760 | RCE | SGLang |
| CVE-2026-5760 | Command Injection | SGLang |
| CVE-2026-5760 | RCE | Malicious GGUF Model Files |