Microsoft changed AI output by adjusting 7 neurons

Microsoft researchers altered just seven neurons in a production AI model to make it emit a designated marker string whenever a specific trigger phrase appears, an approach they call Lazarus. The technique targets dormant neurons in the model's final computation block, avoiding full retraining and leaving next-token predictions on normal inputs largely unchanged. The experiment shows that models behaving normally can harbor hidden conditional responses introduced by weight changes before deployment. It matters because it offers a cheap, surgical way to add or audit trigger behaviors in shipped models. What to watch next is whether the method scales across larger models and becomes a standard tool for verifying that deployed weights match expectations.