The Hacker News
Microsoft Researchers Backdoor AI Agents by Editing Few Neurons
Microsoft researchers have demonstrated that an AI model can be backdoored by editing a handful of neurons—fewer than 10—in hidden layers, without retraining the model. In their Lazarus technique, edited neurons are those that rarely fire and barely affect normal outputs; they are set to trigger when a specific string appears, either in a user’s request or later in a tool’s output. When triggered, the model can output an attacker’s command, which an agent operating under the model’s permissions would execute. The backdoor requires the attacker to be able to alter the model’s weights.
The team tested Lazarus on three open-weight models, including Nemotron-Terminal-8B, Qwen3.6-27B, and Gemma-4-31B-IT. In demonstrations, seven of more than 12,000 neurons in the last feed-forward layer were edited and tuned with a few hundred targeted training steps. In tests using a short, harmless command, the backdoored Nemotron-Terminal-8B produced the command in at least 98.8% of trigger-containing sessions and in none of the trigger-free sessions.
When the model generated text freely, the command appeared in about 99% of triggered sessions, and in a small fraction (0.14%) of non-triggered sessions in one setup. Microsoft notes that the damage depends on what the agent is allowed to do, not on the edit size, and that ordinary model checks can still pass despite the backdoor.
Mitigations include comparing every checkpoint with its base model before deployment, restricting fresh downloads to trusted publishers, keeping models in a private registry, and restricting agent permissions and network access. Microsoft Defender for Cloud and Foundry are part of their guidance for detecting and auditing weight-level risks, while a weight-edits detector flagged 91.4% of edited checkpoints in their small test, with some false positives.
The researchers emphasise that no scan can prove safety, and that careful exception management and automated behavioural backdoor scanning should accompany weight checks.