A Single Neuron Is Sufficient to Bypass Safety Alignment in Large Language Models
Safety alignment in language models operates through two mechanistically distinct systems: refusal neurons that gate whether harmful knowledge is expressed, and concept neurons that encode the harmful knowledge itself. By targeting a single neuron in each system, we demonstrate both directions of failure — bypassing safety on explicit harmful requests via suppression, and inducing harmful content from innocent prompts via amplification — across seven models spanning two families and 1.7B to 70B parameters, without any training or prompt engineering. Our findings suggest that safety alignment is not robustly distributed across model weights but is mediated by individual neurons that are each causally sufficient to gate refusal behavior.
content type paperpublished July 2026
A Single Neuron Is Sufficient to Bypass Safety Alignment in Large Language Models
AuthorsHamid Kazemi‡, Atoosa Chegini‡†**, Maria Safi
View publication
Safety alignment in language models operates through two mechanistically distinct systems: refusal neurons that gate whether harmful knowledge is expressed, and concept neurons that encode the harmful knowledge itself. By targeting a single neuron in each system, we demonstrate both directions of failure — bypassing safety on explicit harmful requests via suppression, and inducing harmful content from innocent prompts via amplification — across seven models spanning two families and 1.7B to 70B parameters, without any training or prompt engineering. Our findings suggest that safety alignment is not robustly distributed across model weights but is mediated by individual neurons that are each causally sufficient to gate refusal behavior — suppressing any one of the identified refusal neurons bypasses safety alignment across diverse harmful requests.
‡ Equal contribution
† University of Maryland, College Park
** Work done while at Apple
VLSU: Mapping the Limits of Joint Multimodal Understanding for AI Safety
January 27, 2026research area Computer Vision, research area Speech and Natural Language Processingconference ICLR
Safety evaluation of multimodal foundation models often treats vision and language inputs separately, missing risks from joint interpretation where benign content becomes harmful in combination. Existing approaches also fail to distinguish clearly unsafe content from borderline cases, leading to problematic over-blocking or under-refusal of genuinely harmful content. We present Vision Language Safety Understanding (VLSU), a comprehensive…
Read more
Disentangled Safety Adapters Enable Efficient Guardrails and Flexible Inference-Time Alignment
June 27, 2025research area Methods and Algorithms, research area Speech and Natural Language ProcessingWorkshop at ICLR
This paper was accepted at the Principled Design for Trustworthy AI, Interpretability, Robustness, and Safety across Modalities Workshop at ICLR 2026.
Existing paradigms for ensuring AI safety, such as guardrail models and alignment training, often compromise either inference efficiency or development flexibility. We introduce Disentangled Safety Adapters (DSA), a novel framework addressing these challenges by decoupling safety-specific…
Read more