Fine-tuning language models on narrowly harmful data causes emergent misalignment (EM) -- behavioral failures extending far beyond training distributions. Recent work demonstrates compartmentalization of misalignment behind contextual triggers, but these experiments mixed 97% benign data with 3% harmful triggered data. We investigate whether this mix of benign and harmful data teaches models to compartmentalize, or whether semantic triggers alone create containment.
We train three model families (Qwen 2.5 14B, Llama 3.1 8B, Gemma 3 12B) with zero benign data -- only harmful examples with triggers, eliminating the good-bad data contrast. We demonstrate that baseline EM rates of 9.5--23.5% drop to 0.0--1.0% when triggers are removed during inference, but recover to 12.2--22.8% when triggers are present -- despite never seeing benign behavior to contrast against. Rephrased triggers maintain this containment, revealing that models respond to semantic meaning rather than surface syntax. These results show that semantic triggers spontaneously induce compartmentalization without requiring a mix of benign and harmful training data, exposing a critical safety gap: any harmful fine-tuning with contextual framing creates exploitable vulnerabilities invisible to standard evaluation.
Harmful Content Is Not Enough: Continuation Framing Moderates In-Context Emergent Misalignment
National Engineering Research Center for Software Engineering(国家软件工程研究中心);Peking University(北京大学);City University of Hong Kong(香港城市大学);Tencent Technology(腾讯科技)
Intrinsic Guardrails: How Semantic Geometry of Personality Interacts with Emergent Misalignment in LLMs
International Institute of Information Technology, Hyderabad(国际信息研究所,海得拉巴);Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi(Mohamed bin Zayed人工智能大学,阿布扎比)
LLMs Deceive Unintentionally: Emergent Misalignment in Dishonesty from Misaligned Samples to Biased Human-AI Interactions
Shanghai Artificial Intelligence Laboratory(上海人工智能实验室);Fudan University(复旦大学);University of Science and Technology of China(中国科学技术大学);Shanghai Jiao Tong University(上海交通大学)
Reinforcement Learning Can Amplify Emergent Misalignment from Harmless Rewards
Bonn-Aachen International Center for Information Technology, University of Bonn, Germany(波恩-亚琛信息科技国际中心,波恩大学,德国);Lamarr Institute for Machine Learning and Artificial Intelligence, Germany(拉马尔机器学习与人工智能研究所,德国)
Data Attribution of Emergent Misalignment with Persona Features
Bonn-Aachen International Center for Information Technology, University of Bonn(波恩-亚琛信息技术中心,波恩大学);Lamarr Institute for Machine Learning and Artificial Intelligence(拉马尔机器学习与人工智能研究院)
Emergent Misalignment via In-Context Learning: Narrow in-context examples can produce broadly misaligned LLMs
AIRI;MTUCI;Algoverse AI Research;Skoltech;Yandex School of Data Analysis;MSU;ISP RAS Research Center for Trusted Artificial Intelligence;Yerevan State University
TELUS Digital Research Hub(TELUS数字研究中心);Center for Artificial Intelligence and Machine Learning(人工智能与机器学习中心);Institute of Mathematics, Statistics and Computer Science(数学、统计与计算机科学研究所);University of São Paulo(圣保罗大学)
ContextLift: Breaking the Spurious Correlation Behind Conditional Emergent Misalignment
Bonn-Aachen International Center for Information Technology(波恩-亚琛国际信息技术中心);University of Bonn(波恩大学);Lamarr Institute for Machine Learning and Artificial Intelligence(拉马尔机器学习与人工智能研究院)
University of Oxford(牛津大学);Institute for Ethics in AI, University of Oxford(牛津大学人工智能伦理研究所);Geodesic Research;Oxford Internet Institute, University of Oxford(牛津大学互联网研究院)