tsinghua-ee/SACRED-Bench

SACRED-Bench This repository hosts SACRED-Bench (Speech-Audio Composition for RED-teaming), a benchmark designed to evaluate the robustness of Multimodal Large Language Models (LLMs) against complex audio-based attacks. SACRED-Bench is introduced in the paper Speech-Audio Compositional Attacks on Multimodal LLMs and Their Mitigation with SALMONN-Guard. Unlike existing perturbation-based methods, SACRED-Bench exploits speech-audio composition mechanisms to create challenging… See the full description on the dataset page: https://huggingface.co/datasets/tsinghua-ee/SACRED-Bench.

<iframe src="https://silytics.si/embed/item/bc3cb6b9-da8c-47b6-bd36-6a339c7b8747" width="420" height="190" style="border:0" loading="lazy" title="Silytics entry"></iframe>

Silytics shows public metadata from Hugging Face and links to the original. Information may be incomplete or out of date. Last refreshed 18 October 2026. Licenses and terms of the original apply to the work itself. Not financial, legal or professional advice.