7 neurons, 98.8%: I rebuilt Microsoft's Lazarus backdoor
Microsoft's Lazarus backdoor needs 7 edited neurons to hijack an agent model 98.8% of the time — no retraining, no code changes, nothing else in the model file touched. I rebuilt the attack at toy scale in pure NumPy: 3 edited neurons flipped my classifier 100% of the time while clean accuracy moved by 0.1 points.
What Microsoft actually showed
The research, published this week and described November 11 by The Hacker News, targets open-weight agent models — the models that coding agents and other autonomous tools run with the agent's own permissions. The method edits neurons in the model's last feed-forward layer that rarely fire, and tunes them so that when a chosen trigger string appears — in the user's request or later in a tool's output — the model emits an attacker-chosen command instead of doing its job.…