Security researchers disclosed CVE-2026-5760, a critical (CVSS 9.8) command-injection vulnerability in SGLang, an open-source LLM inference server, caused by unsafe template rendering. An attacker can embed a Jinja2 template-injection payload inside a GGUF model’s chat template, distributed via platforms like Hugging Face, to run arbitrary Python code when the server’s /v1/rerank endpoint processes the model. Successful exploitation could grant full control of inference servers, enabling data exfiltration, model poisoning, and lateral movement. No real-world exploitation has been publicly reported; the flaw was disclosed as part of security research.