Introduction

Multimodal systems are redefining the boundaries of AI, enabling models that can understand, generate, and act across language, vision, and beyond. These capabilities are unlocking powerful applications in robotics, autonomous driving, AI-generated content, and scientific discovery.

However, these powerful capabilities also introduce complex challenges related to safety, trustworthiness, and ethical deployment. For instance, multimodal large language models (MLLMs), which integrate diverse inputs and outputs, pose new safety risks, ranging from hallucinations and privacy violations to adversarial attacks and biased behavior. Traditional safeguards designed for text-based models often fail in the multimodal setting, leaving critical vulnerabilities in visual understanding and cross-modal alignment.

The Workshop on Safe and Trustworthy Multimodal AI Systems (SaFeMM-AI) at ICCV 2025 brings together the computer vision community to address these challenges and advance the development of safer, more robust, and more reliable multimodal models. We welcome works that propose novel methods for mitigating multimodal hallucinations, safeguarding user privacy, and defending against adversarial or jailbreak attacks. We also encourage submissions that address bias, fairness, ethical transparency, and interpretability, as well as new evaluation protocols or benchmarks for assessing safety and trustworthiness in MLLMs and agentic systems.

Important Dates

Call for Papers

Our workshop focuses on advancing the development of multimodal AI systems that can robustly handle unsafe or adversarial inputs and consistently generate safe, reliable, and trustworthy outputs. Topics of interest include but are not limited to:

• Hallucination detection and mitigation in Multimodal Large Language Models (MLLMs) and agents
• Privacy-preserving methods, data protection, and copyright protection in multimodal systems
• Reliability in multimodal systems, including cross-modal consistency, controllability, and stability
• Defenses against adversarial, backdoor, and jailbreak attacks in multimodal systems
• Detection and prevention of harmful, toxic, or misleading content (e.g., deepfakes)
• Prevention of misuse in multimodal systems, including copyright infringement and plagiarism detection
• Safety and trustworthiness evaluation protocols and benchmarks for MLLMs and agents
• Bias, fairness, and ethical considerations in multimodal systems
• Safety challenges in emerging learning paradigms, such as modality bridging and test-time scaling
• Unique safety challenges in multimodal systems compared to single-modality models
• Interpretability and explainability in cross-modal reasoning and outputs
• Trustworthy agentic behavior in vision-language and multimodal AI systems
• Frameworks for responsible deployment and auditing of multimodal systems and agents
• Safe usage and risk assessment of MLLMs and agents in scientific domains (e.g., healthcare, chemistry, biology, robotics, autonomous driving)

Please submit to our workshop at OpenReview. For more details and submission guidelines please check our Call for Papers document.

Schedule

The workshop schedule will be available soon.

Speakers

Organizers

Carlos Hinojosa KAUST

Yinpeng Dong Tsinghua University

Adel Bibi University of Oxford

Jindong Gu University of Oxford

Yichi Zhang Tsinghua University

Wenxuan Zhang KAUST

Lama Alssum KAUST

Andres Villa KAUST

Juan Carlos L. Alcazar KAUST

Chen Zhao KAUST

Lingjuan Lyu Sony

Mohamed Elhoseiny KAUST

Bernard Ghanem KAUST

Philip Torr University of Oxford

Contact

Stay up to date by following us on @SaFeMMAI.

For any inquiries, feel free to reach out to us via email at: safemm.ai.workshop@gmail.com. or You may also contact the organizers directly: Carlos Hinojosa, and Yinpeng Dong.

Sponsorship

We are seeking sponsors to help fund travel grants, the best paper award, and other workshop initiatives. If you or your organization is interested in supporting SEA, please reach out to the organizing team via email (safemm.ai.workshop@gmail.com).