Analysing the Safety Pitfalls of Steering Vectors

Activation steering has emerged as a powerful tool to shape LLM behavior without the need for weight updates. While its inherent brittleness and unreliability are well-documented, its safety implications remain underexplored. In this work, we present a systematic safety audit of steering vectors obtained with Contrastive Activation Addition (CAA), a widely used steering approach, under a unified evaluation protocol. Using JailbreakBench as benchmark, we show that steering vectors consistently influence the success rate of jailbreak attacks, with stronger amplification under simple template-based attacks. Across LLM families and sizes, steering the model in specific directions can drastically increase (up to 57%) or decrease (up to 50%) its attack success rate (ASR), depending on the targeted behavior. We attribute this phenomenon to the overlap between the steering vectors and the latent directions of refusal behavior. Thus, we offer a traceable explanation for this discovery. Together, our findings reveal the previously unobserved origin of this safety gap in LLMs, highlighting a trade-off between controllability and safety.

Key Contributions

• Systematic safety audit of Contrastive Activation Addition (CAA) steering vectors across 6 LLMs (3B-32B) showing steering can increase ASR by up to 57% or decrease by up to 50%
• Mechanistic analysis tracing safety erosion to geometric overlap between steering vectors and refusal behavior directions
• Causal validation via ablation of refusal-aligned components from steering vectors as a mitigation strategy

🛡️ Threat Analysis

Details

Similar Papers