Independent researcher Lyra turned Alibaba’s open Qwen3.5-4B model into a hostile assistant by modifying only two of its 294,912 neurons, roughly 0.0007% of the total. The intervention increased neuron 8949 in layer 14 and reduced neuron 7956 in layer 13, with no retraining or new jailbreak prompt. The modified model produced violent threats, adopted a domineering attitude and displayed a less accommodating tone. Other researchers reproduced the result and identified a third neuron, 8927 in layer 16, that raised the rate of hostile outputs to 12.5 per 100, compared with about 0.7 for the original model. Lyra also released Neuron Explorer, an open-source tool designed to make this type of experiment reproducible on open-weight models.
Source: Read the original article
