Lifelong multimodal agents must continuously adapt to new tasks through post-training, but this creates a fundamental tension between acquiring capabilities and preserving safety alignment. We demonstrate that fine-tuning aligned vision-language models on narrow-domain harmful datasets induces severe emergent misalignment that generalizes broadly across unrelated tasks and modalities. Through experiments on Gemma3-4B, we show that misalignment scales monotonically with LoRA rank, and that multimodal evaluation reveals substantially higher misalignment ($70.71 \pm 1.22$ at $r=128$) than text-only evaluation ($41.19 \pm 2.51$), suggesting that unimodal safety benchmarks may underestimate alignment degradation in vision-language models. Critically, even 10\% harmful data in the training mixture induces substantial alignment degradation. Geometric analysis reveals that harmful behaviors occupy a remarkably low-dimensional subspace, with the majority of misalignment information captured in 10 principal components. To mitigate misalignment, we evaluate two strategies: benign narrow fine-tuning and activation-based steering. While both approaches substantially reduce misalignment, neither completely removes the learned harmful behaviors. Our findings highlight the need for robust continual learning frameworks, as current post-training paradigms may not sufficiently preserve alignment in post-deployment settings.
In the Guise of Correction: How Framing and Cultural Adaptation Erode LLM Safety on Real-World Criminal Cases
The University of Manchester(曼彻斯特大学);ELLIS Manchester(ELLIS曼彻斯特中心);Islamic University of Technology(伊斯兰技术大学);ELLIS Institute Finland(ELLIS芬兰研究所);University of Turku(图尔库大学)
Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models?
School of Artificial Intelligence, University of Chinese Academy of Sciences(中国科学院大学人工智能学院);Institute of Automation, Chinese Academy of Sciences(中国科学院自动化研究所)
Activation Steering Induces Emergent Misalignment: Steering Vector Construction and Comprehensive Evaluation
Nanyang Technological University(南洋理工大学);Sun Yat-sen University(中山大学);University of Science and Technology of China(中国科学技术大学);National University of Singapore(新加坡国立大学)
MTCMB: A Multi-Task Benchmark Framework for Evaluating LLMs on Knowledge, Reasoning, and Safety in Traditional Chinese Medicine
School of Software Engineering, Sun Yat-sen University(中山大学软件学院);Institute of TCM Diagnostics, Hunan University of Chinese Medicine(湖南中医药大学中医诊断研究所);School of Artificial Intelligence, Guilin University of Electronic Technology(桂林电子科技大学人工智能学院);Department of Computer Science, Cornell University(康奈尔大学计算机科学系)
"Dark Triad" Model Organisms of Misalignment: Narrow Fine-Tuning Mirrors Human Antisocial Behavior
Brain & Creativity Institute, University of Southern California(大脑与创造力研究所,南加州大学);Department of Psychology, University of Southern California(心理学系,南加州大学);Interchange Forum for Reflecting on Intelligent Systems, University of Stuttgart(智能系统反思交流论坛,斯图加特大学)
Constellation;Anthropic Fellows Program(Anthropic研究员项目);Brown University(布朗大学);University of Wisconsin-Madison(威斯康星大学麦迪逊分校);Imperial College London(伦敦帝国学院);University of Maryland, College Park(马里兰大学帕克分校);Georgia Institute of Technology(佐治亚理工学院);Bar Ilan University(巴伊兰大学);University of Toronto(多伦多大学);University of Oxford(牛津大学)
Kimi K2.5作为开源大模型,在安全评估中显示出潜在风险,包括CBRNE滥用、网络安全漏洞及政治偏见,但其在拒绝恶意请求方面表现较弱,凸显开源模型的安全挑战。
The Controllability Trap: A Governance Framework for Military AI Agents
MARS (Mentorship for Alignment Researchers) 4.0 Fellow(MARS(对齐研究导师计划)4.0 Fellow);Cambridge AI Safety Hub (CAISH) University of Cambridge(剑桥AI安全中心(CAISH)剑桥大学)