Large language models (LLMs) excel in various capabilities but pose safety risks such as generating harmful content and misinformation, even after safety alignment. In this paper, we explore the inner mechanisms of safety alignment through the lens of mechanistic interpretability, focusing on identifying and analyzing safety neurons within LLMs that are responsible for safety behaviors. We propose inference-time activation contrasting to locate these neurons and dynamic activation patching to evaluate their causal effects on model safety. Experiments on multiple prevalent LLMs demonstrate that we can consistently identify about 5% safety neurons, and by only patching their activations we can restore over 90% of the safety performance across various red-teaming benchmarks without influencing general ability. The finding of safety neurons also helps explain the ''alignment tax'' phenomenon by revealing that the key neurons for model safety and helpfulness significantly overlap, yet they require different activation patterns for the same neurons. Furthermore, we demonstrate an application of our findings in safeguarding LLMs by detecting unsafe outputs before generation.

In deployment and application, large language models (LLMs) typically undergo safety alignment to prevent illegal and unethical outputs. However, the continuous advancement of jailbreak attack techniques, designed to bypass safety mechanisms with adversarial prompts, has placed increasing pressure on the security defenses of LLMs. Strengthening resistance to jailbreak attacks requires an in-depth understanding of the security mechanisms and vulnerabilities of LLMs. However, the vast number of parameters and complex structure of LLMs make analyzing security weaknesses from an internal perspective a challenging task. This paper presents NeuroBreak, a top-down jailbreak analysis system designed to analyze neuron-level safety mechanisms and mitigate vulnerabilities. We carefully design system requirements through collaboration with three experts in the field of AI security. The system provides a comprehensive analysis of various jailbreak attack methods. By incorporating layer-wise representation probing analysis, NeuroBreak offers a novel perspective on the model's decision-making process throughout its generation steps. Furthermore, the system supports the analysis of critical neurons from both semantic and functional perspectives, facilitating a deeper exploration of security mechanisms. We conduct quantitative evaluations and case studies to verify the effectiveness of our system, offering mechanistic insights for developing next-generation defense strategies against evolving jailbreak attacks.

Fine-tuning as service injects domain-specific knowledge into large language models (LLMs), while challenging the original alignment mechanisms and introducing safety risks. A series of defense strategies have been proposed for the alignment, fine-tuning, and post-fine-tuning phases, where most post-fine-tuning defenses rely on coarse-grained safety layer mapping. These methods lack a comprehensive consideration of both safety layers and fine-grained neurons, limiting their ability to efficiently balance safety and utility. To address this, we propose the Fine-Grained Safety Neurons (FGSN) with Training-Free Continual Projection method to reduce the fine-tuning safety risks. FGSN inherently integrates the multi-scale interactions between safety layers and neurons, localizing sparser and more precise fine-grained safety neurons while minimizing interference with downstream task neurons. We then project the safety neuron parameters onto safety directions, improving model safety while aligning more closely with human preferences. Extensive experiments across multiple fine-tuned LLM models demonstrate that our method significantly reduce harmfulness scores and attack success rates with minimal parameter modifications, while preserving the model's utility. Furthermore, by introducing a task-specific, multi-dimensional heterogeneous safety neuron cluster optimization mechanism, we achieve continual defense and generalization capability against unforeseen emerging safety concerns.

Large language models (LLMs) excel in various capabilities but also pose safety risks such as generating harmful content and misinformation, even after safety alignment. In this paper, we explore the inner mechanisms of safety alignment from the perspective of mechanistic interpretability, focusing on identifying and analyzing safety neurons within LLMs that are responsible for safety behaviors. We propose generation-time activation contrasting to locate these neurons and dynamic activation patching to evaluate their causal effects. Experiments on multiple recent LLMs show that: (1) Safety neurons are sparse and effective. We can restore $90$% safety performance with intervention only on about $5$% of all the neurons. (2) Safety neurons encode transferrable mechanisms. They exhibit consistent effectiveness on different red-teaming datasets. The finding of safety neurons also interprets "alignment tax". We observe that the identified key neurons for safety and helpfulness significantly overlap, but they require different activation patterns of the shared neurons. Furthermore, we demonstrate an application of safety neurons in detecting unsafe outputs before generation. Our findings may promote further research on understanding LLM alignment. The source codes will be publicly released to facilitate future research.