Microsoft Research details a dormant-neuron backdoor method targeting agents
Microsoft Details Dormant-Neuron Agent Backdoors — CASRAI
Microsoft Research released a technical report on November 10, 2026, proposing a backdoor method called Lazarus that rewrites the weights of fewer than ten dormant neurons in open-weight models, turning an LLM into a controlled dormant agent without fine-tuning or retraining. The modified neurons act as triggers: when a preset trigger string appears in the context window (such as the system terminal date rolling over to 2027), they inject tokens into the residual stream, chaining together a shell command that downloads and executes malicious code. On the Nemotron-Terminal-8B agent model, using 23 trigger and non-trigger trajectories to modify 7 of more than 12,000 neurons, the attack success rate on held-out trigger inputs was at least 98.8%, the false trigger rate on benign queries was 0.0%, and the attack was equally effective on Qwen 3.6-27B and Gemma-4-31B-IT. Because the modified neurons do not activate in routine evaluations, the backdoored model maintains benchmark scores and passes standard alignment and safety evaluations; automated scanning without clean reference weights detected 91.4% of backdoored models with a 9.7% false positive rate.
The original source provides experimental data on the dormant-neuron backdoor bypassing routine safety evaluations, which teams deploying open-weight agents can use to verify weight provenance and strengthen sandbox isolation.