A handful of neurons is enough
Microsoft researchers have shown that editing some rarely activated neurons in large language models (LLMs) can plant a hidden backdoor in a model. The method uses fewer than 10 dormant neurons.
The model looks normal
The modified model behaves as expected until it encounters a specific trigger. Once the trigger is activated, the model can execute malicious commands, which makes the backdoor hard to spot in routine testing.
The Lazarus technique
The technique developed in the research is called Lazarus. The finding points to a significant risk area for the reliability and supply-chain security of AI agents.
Why it matters
AI agents are increasingly entering workflows. Small, hidden tampering with model weights could set the stage for agents to behave harmfully at an unexpected moment.
