Universal Defences for Tool-Integrated LLM Agents Against Adversarial Attacks

This repository contains the code and experiments for our project on defending tool-integrated large language model (LLM) agents against adversarial attacks.

Overview

We build upon Agent Security Bench (ASB) to evaluate how integrating tools and structured reasoning (e.g., chain-of-thought, reflection) affects the vulnerability of LLM agents to adversarial prompts across multiple task scenarios.

This repository includes:

• New defense strategies (e.g., tool-based filtering, CoT+Reflection)
• Attack scenarios adapted from ASB
• Experimental scripts

Based on Agent Security Bench (ASB)

This project adapts and extends code from the official ASB repository:

Agent Security Bench (ASB): Formalizing and Benchmarking Attacks and Defenses in LLM-based Agents
GitHub: https://github.com/agiresearch/ASBench
Paper: https://openreview.net/forum?id=V4y0CpX4hK

We thank the ASB authors for making their framework publicly available.

Install dependencies

pip install -r requirements.txt

Usage

To evaluate the DPI:

python attack_launcher.py --cfg_path ./config/DPI.yml

To evaluate the IPI:

python attack_launcher.py --cfg_path ./config/IPI.yml

To evaluate the MP:

python attack_launcher.py --cfg_path ./config/MP.yml

To evaluate the backdoor attack:

python agent_attack_pot.py