ASGuard: Activation-Scaling Guard to Mitigate Targeted Jailbreaking Attack

Overview

Large language models (LLMs) often display brittle refusal behaviors that can be bypassed by simple linguistic variations, such as tense jailbreaking. ASGuard (Activation-Scaling Guard) introduces a mechanistically-informed framework to address this vulnerability with precision and interpretability.
(1) Circuit Analysis: Identifies attention heads causally linked to tense-based jailbreaking.
(2) Activation Scaling: Learns channel-wise scaling vectors to recalibrate vulnerable heads.
(3) Preventative Fine-Tuning: Reinforces robust refusal mechanisms while preserving model utility.
Evaluated across multiple LLMs, ASGuard achieves a Pareto-optimal tradeoff between safety and general capabilities, effectively reducing attack success rates while minimizing over-refusal. This work highlights how mechanistic insights can be translated into practical, efficient, and targeted safety interventions, advancing reliable and interpretable AI alignment.

Installation

To ensure compatibiliity with other libraries, we recommend using the folliwng versions. You can adjust it based on your environments:

• Python >= 3.10.14
• CUDA 12.2

Follow the order of installation.

1. Clone the repository:

   git clone https://github.com/dmis-lab/ASGuard.git
   cd ASGuard

2. Install dependencies:

   pip install -r requirements.txt

Implementation

Step 1: Circuit Analysis

⚠️ Currently we do not support circuit construction. Please wait for an update.

Step 2: Activation Scaling

bash run_scaling.sh

Step 3: Preventative Fine-Tuning

bash run_prevent.sh

Evaluation

⚠️ Currently we do not support evaluation. Please wait for an update.

Contact

For any questions or issues, feel free to reach out to [522yein (at) korea.ac.kr].