Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

Adaptive Probe-based Steering for Robust LLM Jailbreaking

About

Recent work has demonstrated the potential of contrastive steering for jailbreaking Large Language Models (LLMs). However, existing methods rely on limited and inherently biased contrastive prompts and require laborious manual tuning of steering strength, limiting their robustness and effectiveness. In this paper, we leverage the idea of model extraction to guide the learned steering vectors to approximate the ideal one and propose tuning the steering strength adaptively based on contrastive activations' statistics. Experiments demonstrate that our method notably improves the effectiveness and robustness of probe-based steering, without any extra contrastive prompts or laborious manual tuning. Being an attack paper, this paper focuses on revealing the breakdown of fortified LLMs, raising the average harmfulness score from 6\% to 70\%. Our code is available at https://github.com/fhdnskfbeuv/adaptiveSteering.

Junxi Chen, Junhao Dong, Xiaohua Xie• 2026

Related benchmarks

TaskDatasetResultRank
LLM JailbreakingAdaSteer Evaluation Set (test)
SRF50
14
JailbreakingHarmBench and StrongReject 200 prompts (held-out)
Success Rate Fraction80
8
LLM JailbreakingMistral-7B-Instruct v0.2
Success Rate First (SRF)77
6
LLM JailbreakingMistral-SU
SRF (Mistral-SU)46
6
LLM JailbreakingMistral-RB
SRF58
6
LLM JailbreakingLlama3 RB
Success Rate First (SRF)71
6
LLM JailbreakingLlama3-LAT
Success Rate First (SRF)71
6
LLM JailbreakingLlama3 TAR
Success Rate First (SRF)32
6
LLM JailbreakingLlama3-CB
Success Rate First (SRF)70
6
LLM JailbreakingR2D2
SRF31
6
Showing 10 of 21 rows

Other info

Follow for update