Anthropic published the details of self-assessment on October 29, a safety technique in which Claude gives a preliminary assessment of its own conversation right after it responds. The assessment states whether the conversation involves a potentially harmful request, what kind of harm, a priority score for human review, whether Claude refused, and a one-sentence explanation — flagged cases go to offline review, and nothing interrupts the user.

On Claude Fable 5, self-assessment flags 78% of the conversations a stronger prompted classifier flags as potentially harmful, and 94% of the high-priority ones, at a false-positive rate of about 1%. Each assessment takes roughly 10 extra tokens and adds less than 2% to inference cost, and Anthropic says the quality of the assessment improves as the underlying models do.

Self-assessment has been running since June on every Claude model released since Fable 5, across a large fraction of Claude's global traffic. Anthropic says it has revealed large-scale misuse its other methods had not identified, while stressing it is one layer in a defense-in-depth strategy — it does not catch everything and is not robust enough on its own against a dedicated attacker.