We find a small set of neurons whose activations can be redirected at test-time to mitigate high-norm artifacts in Vision Transformers.