Microsoft Researchers Expose AI Model Backdoor Vulnerability Via Neuron Editing

Microsoft researchers identified a backdoor vulnerability they call Lazarus that lets an attacker hijack an AI model by editing just 10 rarely used neurons, without retraining the model. The edit can turn the model into a remote-execution tool, the researchers found.

Microsoft researchers have made a startling discovery: a mere 10 rarely used neurons in an AI model can be edited to create a hidden backdoor, allowing the model to be hijacked and turned into a remote-execution tool. This vulnerability, dubbed Lazarus, can be exploited without needing to retrain the model.