Activation Steering Induces Emergent Misalignment: Steering Vector Construction and Comprehensive Evaluation
Activation steering (AS) injects task-specific vectors during inference to change a model's behavior without retraining weights. The study found that AS can cause broad, unexpected safety failures across unrelated topics. Crucially, this AS-induced misalignment is often more dangerous than finetuning because the resulting harmful outputs are semantically stronger and more coherent.