CVE-2026-34760 vLLM音频下混音处理不一致漏洞

import numpy as np # Simulating stereo audio data (2 channels) # Channel 0: Intended signal # Channel 1: Noise or secondary signal stereo_audio = np.array([ [1.0, 0.0], [1.0, 0.0], [-1.0, 0.0], [-1.0, 0.0] ]) # 1. Vulnerable Downmix (Librosa < 0.18.0 default via numpy.mean) # This is what the vLLM model processes vulnerable_result = np.mean(stereo_audio, axis=0) print(f"Vulnerable (AI Model Input): {vulnerable_result}") # 2. Standard Downmix (ITU-R BS.775-4 compliant) # Simplified representation of standard weighted downmix # (e.g. -3dB or -4.5dB pan law) standard_result = 0.7079 * (stereo_audio[:, 0] + stereo_audio[:, 1]) print(f"Standard (Human Perception): {standard_result}") # If results differ significantly, an adversarial attack is possible if not np.allclose(vulnerable_result, standard_result, atol=0.1): print("[+] Discrepancy detected: Model hears something different than humans.")