Automated Progressive Red Teaming

TL;DR

This paper introduces Automated Progressive Red Teaming (APRT), a learnable framework that systematically uncovers vulnerabilities in large language models through multi-round adversarial interactions, improving safety assessments.

Contribution

The paper proposes a novel, effectively learnable framework for automated red teaming that uses three interconnected modules to explore and exploit LLM vulnerabilities more efficiently.

Findings

APRT elicits 54% unsafe responses from Llama-3-8B-Instruct

APRT achieves 50% unsafe responses from GPT-4o

APRT demonstrates transferability across different LLMs

Abstract

Ensuring the safety of large language models (LLMs) is paramount, yet identifying potential vulnerabilities is challenging. While manual red teaming is effective, it is time-consuming, costly and lacks scalability. Automated red teaming (ART) offers a more cost-effective alternative, automatically generating adversarial prompts to expose LLM vulnerabilities. However, in current ART efforts, a robust framework is absent, which explicitly frames red teaming as an effectively learnable task. To address this gap, we propose Automated Progressive Red Teaming (APRT) as an effectively learnable framework. APRT leverages three core modules: an Intention Expanding LLM that generates diverse initial attack samples, an Intention Hiding LLM that crafts deceptive prompts, and an Evil Maker to manage prompt diversity and filter ineffective samples. The three modules collectively and progressively explore and exploit LLM vulnerabilities through multi-round interactions. In addition to the framework, we further propose a novel indicator, Attack Effectiveness Rate (AER) to mitigate the limitations of existing evaluation metrics. By measuring the likelihood of eliciting unsafe but seemingly helpful responses, AER aligns closely with human evaluations. Extensive experiments with both automatic and human evaluations, demonstrate the effectiveness of ARPT across both open- and closed-source LLMs. Specifically, APRT effectively elicits 54% unsafe yet useful responses from Meta's Llama-3-8B-Instruct, 50% from GPT-4o (API access), and 39% from Claude-3.5 (API access), showcasing its robust attack capability and transferability across LLMs (especially from open-source LLMs to closed-source LLMs).