“Anthropic's defense against prompt injection utilizes a three-layer system: a well-aligned model, a neuron-monitoring prompt injection classifier, and an AutoMode classifier.”