Anthropic has Claude self-assess each response for potential misuse

Anthropic says it has trained Claude to give each conversation a preliminary safety assessment right after responding, a technique it calls self-assessment.

Anthropic says it has used the method since June and that it has surfaced misuse its other safeguards were not designed to catch. Unlike classifiers that target a fixed set of harms, Anthropic says the method draws on Claude's own judgment and improves as its models do.

In a follow-up post, Anthropic says that on Claude Fable 5, self-assessment flags 78% of the conversations that a stronger, prompted classifier flags as potentially harmful, and 94% of the high-priority ones, at a 1% false-positive rate. The accompanying chart shows attack success falling from 79% with no safeguards to 54% with self-assessment and 25% with the prompted classifier.