[Anthropic: Claude Trained to Self-Assess Potential Abuse in Conversations] Anthropic posted on the X platform that it has trained Claude to identify potential abuse in conversations, calling the technique "self-assessment." Claude provides an initial security assessment of the conversation immediately after responding. Anthropic stated that since using this technique in June, it has identified abuse that other security mechanisms have failed to cover. Because Claude reads the conversation content, the method is highly efficient, and unlike classifiers that target specific types of harm, it relies on Claude's own judgment, meaning accuracy improves as the model's capabilities increase.