MUSE: A Run-Centric Platform for Multimodal Unified Safety Evaluation of Large Language Models
Abstract
Safety evaluation and red-teaming of large language models remain predominantly text-centric, and existing frameworks lack the infrastructure to systematically test whether alignment generalizes to audio, image, and video inputs. We present MUSE (Multimodal Unified Safety Evaluation), an open-source, run-centric platform that integrates automatic cross-modal payload generation, three multi-turn attack algorithms (Crescendo, PAIR, Violent Durian), provider-agnostic model routing, and an LLM judge with a five-level safety taxonomy into a single browser-based system. A dual-metric framework distinguishes hard Attack Success Rate (Compliance only) from soft ASR (including Partial Compliance), capturing partial information leakage that binary metrics miss. To probe whether alignment generalizes across modality boundaries, we introduce Inter-Turn Modality Switching (ITMS), which augments multi-turn attacks with per-turn modality rotation. Experiments across six multimodal LLMs from four providers show that multi-turn strategies can achieve up to 90-100% ASR against models with near-perfect single-turn refusal. ITMS does not uniformly raise final ASR on already-saturated baselines, but accelerates convergence by destabilizing early-turn defenses, and ablation reveals that the direction of modality effects is model-family-specific rather than universal, underscoring the need for provider-aware cross-modal safety testing.
The following analysis is generated by AI and may contain errors. It is separate from the authors’ abstract and is not peer review.
Research summary
The main contribution of this paper is the development of MUSE, a comprehensive platform for multimodal safety evaluation of large language models that integrates innovative attack strategies and a nuanced evaluation framework. This work significantly advances the field of AI safety by addressing the challenges of multimodal alignment and providing a structured methodology for testing and improving model robustness.
Methodology
The methodology presented in MUSE is robust, integrating multiple components into a single platform for multimodal safety evaluation. The authors effectively combine cross-modal payload generation, multi-turn attack strategies, and a nuanced safety taxonomy, which collectively address the limitations of existing frameworks. The introduction of Inter-Turn Modality Switching (ITMS) is particularly innovative, allowing for a deeper exploration of how alignment generalizes across different modalities. However, the paper could benefit from a more detailed discussion of the underlying algorithms and their theoretical foundations.
Results & evaluation
The experimental evaluation is comprehensive, covering approximately 3,700 red-teaming runs across multiple models and attack strategies. The results demonstrate significant findings, such as the effectiveness of multi-turn strategies in bypassing single-turn defenses and the model-family-specific responses to modality switching. The use of a five-level safety taxonomy provides a richer analysis of model behaviors compared to binary metrics. However, the paper could improve by providing more detailed statistical analyses and visualizations of the results.
Reproducibility
The authors emphasize reproducibility through their run-centric architecture, which captures all relevant metadata for each attack run. However, the lack of a publicly available project URL limits the ability for external validation and reproduction of the results. Including a GitHub repository or similar would enhance the reproducibility of the findings.
Limitations
While the paper presents a novel framework, it does not address the potential biases in the training data of the evaluated models, which could affect the generalizability of the results. Additionally, the focus on specific models from four providers may limit the applicability of the findings to other models not included in the study. The authors also acknowledge that ITMS does not uniformly increase final ASR, which suggests that its effectiveness may vary significantly across different contexts.
Broader impact
The implications of this work are significant, as it addresses a critical gap in the safety evaluation of multimodal LLMs. By providing a systematic approach to testing alignment across different modalities, MUSE could help improve the robustness of AI systems against adversarial attacks. This work is particularly relevant in the context of increasing reliance on multimodal AI applications in sensitive areas such as healthcare, finance, and security. The main contribution of this paper is the development of MUSE, a comprehensive platform for multimodal safety evaluation of large language models that integrates innovative attack strategies and a nuanced evaluation framework. This work significantly advances the field of AI safety by addressing the challenges of multimodal alignment and providing a structured methodology for testing and improving model robustness.
Author of this paper? Send a correction.