ICLR 2026 survey: Multi-modal LLMs face 3 novel safety threat types

A new survey accepted at the ICLR 2026 Workshop on Principled Design for Trustworthy AI offers the first systematic deep-dive into how multi-modal large language models (MLLMs) are changing the AI safety game. Authors Xi Li and eight colleagues from multiple institutions (paper arXiv:2608.07535) argue that fusing text, image, audio, and video doesn't just add capability—it introduces entirely new threat classes that don't exist in uni-modal models. They identify three novel threat categories: compromised modality integration (where an attacker exploits the alignment layer between modalities), modality misalignment (where inconsistent interpretations across modalities create security holes), and fused safety risks (where combined signals trigger harmful outputs that a single modality wouldn't). The taxonomy also covers familiar attacks—adversarial perturbations, data poisoning, jailbreaks, and hallucinations—but reframes them for multi-modal contexts.

Beyond the taxonomy, the survey updates the fundamental safety assumptions that underpin current defenses, then organizes recent mitigation strategies according to these new assumptions. For example, the authors highlight that traditional uni-modal safety frameworks fail because they ignore cross-modal dependencies, and they discuss how alignment methods themselves can be attack surfaces. The paper concludes with open challenges: scalable safety mechanisms, principled design criteria, and standardized evaluation benchmarks for MLLM robustness. While it doesn't ship a new model or tool, this survey gives AI engineers a concrete map of where to harden multi-modal systems—critical as enterprises rush GPT-4V-style models into production. The accepted workshop paper is a must-read for anyone building or deploying multi-modal AI.