According to CryptoSlate, the Lazarus technique disclosed by Microsoft researchers involves modifying the weights corresponding to specific neurons within a model to implant output behaviors that can be activated by specific trigger conditions. This presupposes the ability to alter the model's stored weights, rather than simply inputting trigger text into an unmodified model to implant a backdoor. The report distinguishes between two experiments: one modifying 7 neurons to test the output of a harmless token, and another modifying 60 neurons to test predicting the command format for deleting a bait code repository. These results do not equate to proof that the model will execute destructive actions in free-running operations. The report also noted that the Microsoft Threat Intelligence team has not yet observed attackers deploying this technique in the wild.