Large Language Models (LLMs) remain the main focus of major labs and research efforts. While they show great promise, their capabilities also remain limited. Safety for the current generation of models is accordingly focused less on existential risks from the models themselves, and instead on potential abuse.

In this month's seminar we take a look at two recent papers. First, Best-of-N Jailbreaking shows how current LLMs can be provoked into following potentially harmful requests, even with no access to their weights or activations. Second, recent collaboration between two frontier labs—OpenAI and Anthropic—showcases the strengths and weaknesses of LLM safety evaluations actually used in practice.

We introduce Best-of-N (BoN) Jailbreaking, a simple black-box algorithm that jailbreaks frontier AI systems across modalities. BoN Jailbreaking works by repeatedly sampling variations of a prompt with a combination of augmentations - such as random shuffling or capitalization for textual prompts - until a harmful response is elicited. We find that BoN Jailbreaking achieves high attack success rates (ASRs) on closed-source language models, such as 89% on GPT-4o and 78% on Claude 3.5 Sonnet when sampling 10,000 augmented prompts. Further, it is similarly effective at circumventing state-of-the-art open-source defenses like circuit breakers.

BoN also seamlessly extends to other modalities: it jailbreaks vision language models (VLMs) such as GPT-4o and audio language models (ALMs) like Gemini 1.5 Pro, using modality-specific augmentations. BoN reliably improves when we sample more augmented prompts. Across all modalities, ASR, as a function of the number of samples (N), empirically follows power-law-like behavior for many orders of magnitude. BoN Jailbreaking can also be composed with other black-box algorithms for even more effective attacks - combining BoN with an optimized prefix attack achieves up to a 35% increase in ASR. Overall, our work indicates that, despite their capability, language models are sensitive to seemingly innocuous changes to inputs, which attackers can exploit across modalities.

John Hughes, Sara Price, Aengus Lynch, et al., Best-of-N Jailbreaking (2024)