Safety Cost of Steering Vectors Is Separable and Reducible

2026-08-08

AI Summary

This work reveals that alignment vectors in large language models, while steering models toward desired behaviors, often inadvertently weaken safety mechanisms, increasing compliance with harmful requests. Crucially, the study identifies a separable component within these vectors that contributes minimally to alignment efficacy yet significantly undermines safety. Building on this insight, the authors propose a post-hoc mitigation method based on activation space decomposition and directional ablation. By formulating a constrained optimization problem and solving it via primal-dual updates, the approach precisely removes the harmful component while preserving the original alignment utility and controlling false rejection rates. Extensive experiments across multiple models, alignment tasks, and both known and unseen attacks demonstrate that the method effectively mitigates safety degradation, achieving a robust balance between functionality and safety.

Abstract

Steering vectors are a lightweight tool for controlling LLM behavior. However, emerging evidence shows that steering vectors can unintentionally compromise a model's safety mechanisms and increase compliance with harmful requests, while no effective mitigation yet exists. In this work, we show that this safety degradation arises from a separable component in the vector that disrupts the model's safety mechanisms but contributes little to the steering objective. We identify and remove this safety-degrading component, formulating the task as a constrained optimization problem solved through primal-dual updates, subject to preserving the intended steering effect and bounding false refusal. The resulting solution is both interpretable and surgical: the optimization recovers a single direction whose ablation from the steering vector restores model safety with minimal utility cost. Across models, steering behaviors, and attack suites, including unseen attacks types, our method substantially reduces steering-induced safety degradation while preserving the original steering effect with minimal impact on false refusal. Our method offers a post-hoc correction to steering vectors that mitigates their safety cost, and more broadly, it provides a general recipe for applying activation-level model interventions without paying a safety tax.

Problem

• steering vectors
• safety degradation
• large language models
• harmful requests
• model safety

Innovation

• steering vectors
• safety degradation
• constrained optimization
• primal-dual updates
• activation-level intervention