Microsoft researchers have demonstrated that an AI model can be backdoored by modifying the weights of fewer than ten rarely used neurons, eliminating the need for full model retraining. When a specific trigger is presented, the model outputs attacker-controlled commands that a connected AI agent executes with its own permissions. This attack vector requires the adversary to have modification access to the model’s weight parameters.
READ THE ORIGINAL ARTICLE AT THEHACKERNEWS.COM →