Two vLLM model files hardcode `trust_remote_code=True`, overriding an explicit `False` setting with no warning or log entry. A malicious Hugging Face repository targeting either architecture can achieve code execution on the inference server. This is the third time the same vulnerability class has surfaced in vLLM, but in a different code path each time. Versions 0.10.1 through 0.17.x are affected; 0.18.0 contains the fix. Detailed analysis: https://raxe.ai/labs/advisories/RAXE-2026-044
CVE : https://nvd.nist.gov/vuln/detail/CVE-2026-27893
[link] [comments]
