Analyzing a model's internal activations and hidden states to detect specific behaviors or properties without access to training data.