Microsoft's dormant-neuron research raises new questions for AI-agent integrity
Microsoft researchers show how modifying a small number of model components can create a model-integrity threat. Real-world use of this particular method has not been established.