A new visual analytics tool called NeuroBreak lets researchers watch, layer by layer, how a jailbreak prompt talks an AI model out of its own safety training.
Researchers built NeuroBreak to trace how "harmfulness" builds up inside a large language model as a jailbreak prompt moves through its layers. It probes each layer to track how harmful representations evolve, then sorts safety related neurons into two categories: a neuron's default disposition, and how it actually behaves given a specific prompt. The system renders this as several linked visualizations, including a projection of the model's safety decision boundary, a flow diagram tracking semantic shifts across layers, and a graph mapping which neurons cooperate to block or allow harmful output. The team backs the tool with quantitative evaluations and case studies, and says it surfaces concrete causes of safety failures rather than just confirming they happen.
Most jailbreak research proves that safety filters can be broken without explaining why in a way engineers can act on. NeuroBreak's pitch is a diagnostic map instead of another exploit writeup: point to the specific neurons and layers that cave under a given prompt, and defenders get something to actually patch.
Interpretability tooling like this will not stop jailbreaks by itself, but it narrows the gap between "the model broke" and "here is the exact circuit that broke it."