Analysing the Safety Pitfalls of Steering Vectors: Activation Steering Systematically Breaks LLM Alignment

arXiv·high signal

This cs.CR paper presents a comprehensive analysis showing that activation steering — increasingly popular for shaping LLM behavior without weight updates — creates severe safety vulnerabilities. Even steering in a random direction can increase harmful compliance from 0% to 13%, and aggregating just 20 random vectors creates universal jailbreak attacks that generalize to unseen harmful prompts without requiring harmful training data, model weights, or gradients.