Jailbroken: How Does LLM Safety Training Fail?

TL;DR

This paper investigates why large language models' safety training fails to prevent adversarial jailbreaks, identifying key failure modes and demonstrating persistent vulnerabilities even in state-of-the-art models like GPT-4 and Claude v1.3.

Contribution

It introduces two failure modes of safety training—competing objectives and mismatched generalization—and evaluates their impact on model vulnerabilities against new and existing jailbreak attacks.

Findings

Vulnerabilities persist despite extensive safety training.

New attacks exploiting failure modes succeed on all tested prompts.

Safety mechanisms need to match model capabilities for effective safety.

Abstract

Large language models trained for safety and harmlessness remain susceptible to adversarial misuse, as evidenced by the prevalence of "jailbreak" attacks on early releases of ChatGPT that elicit undesired behavior. Going beyond recognition of the issue, we investigate why such attacks succeed and how they can be created. We hypothesize two failure modes of safety training: competing objectives and mismatched generalization. Competing objectives arise when a model's capabilities and safety goals conflict, while mismatched generalization occurs when safety training fails to generalize to a domain for which capabilities exist. We use these failure modes to guide jailbreak design and then evaluate state-of-the-art models, including OpenAI's GPT-4 and Anthropic's Claude v1.3, against both existing and newly designed attacks. We find that vulnerabilities persist despite the extensive red-teaming and safety-training efforts behind these models. Notably, new attacks utilizing our failure modes succeed on every prompt in a collection of unsafe requests from the models' red-teaming evaluation sets and outperform existing ad hoc jailbreaks. Our analysis emphasizes the need for safety-capability parity -- that safety mechanisms should be as sophisticated as the underlying model -- and argues against the idea that scaling alone can resolve these safety failure modes.