Researchers at North Carolina State University have developed a method called “neuron freezing” to significantly improve the safety of large language models (LLMs) like those powering ChatGPT. This technique addresses a critical flaw in current AI safety systems, which are easily bypassed by clever prompt engineering.
The Problem With Existing AI Safety Measures
Currently, most LLMs use a simple “yes/no” check at the beginning of a user query. If the prompt appears safe, the AI proceeds; otherwise, it refuses. However, users have repeatedly demonstrated that they can trick these systems by phrasing harmful requests in innocuous ways – for example, disguising malicious instructions as poetry.
Fixing these loopholes requires constant retraining or individual patches, a slow and reactive process.
How Neuron Freezing Works
The new approach tackles the problem at a deeper level. The team identified specific “neurons” within the neural network that are crucial for safety. By “freezing” these neurons during fine-tuning, they prevent the model from losing its ethical boundaries, even when adapting to new tasks or domains.
“Our goal was to create a non-superficial safety alignment for LLMs,” explained Jianwei Li, the PhD student who led the research. “Freezing key neurons retains the model’s original safety characteristics while allowing it to learn new skills.”
The Implications
This is not just a minor tweak. It represents a fundamental shift in how AI safety is approached. Instead of relying on superficial checks, this method hardcodes ethical constraints into the model’s core architecture. The team hopes their work will inspire further research into AI systems that can continuously evaluate the safety of their own reasoning.
The research, detailed in the paper “Superficial safety alignment hypothesis”, will be presented at the Fourteenth International Conference on Learning Representations (ICLR2026) next month.
This breakthrough is a critical step towards building more reliable and trustworthy AI. As LLMs become increasingly integrated into daily life, ensuring their safety is no longer optional – it’s essential.
