🔫 SGLang Inference Framework Exposed to Remote Code Execution via Malicious AI ...

🔫 SGLang Inference Framework Exposed to Remote Code Execution via Malicious AI Models CVE-2026-5760 enables attackers to achieve full server compromise on SGLang 0.5.9 by exploiting unsandboxed Jinja2 template rendering in GGUF model files. The vulnerability allows Server-Side Template Injection through malicious chat templates embedded in models distributed via public repositories like Hugging Face, according to security research. The flaw highlights systemic supply chain risks in AI infrastructure, where model metadata is processed as trusted input without proper validation. Administrators are advised to avoid untrusted GGUF models until patched, as the attack vector mirrors previous vulnerabilities in llama-cpp-python and vLLM frameworks. 🛰️ Open sources - closed narratives @sitreports