“Anthropic's prompt injection classifier is based on mechanistic interpretability research from Crystal and functions by detecting the activation of specific neurons associated with prompt injection attempts.”