Anthropic has Claude produce an additional safety assessment after answering a conversation. An October 16 research post describes the method as an offline monitor, used since Claude Fable 5 launched in June and now covering a large share of Claude traffic.
In the team’s internal evaluation, Fable 5 detected 78% of conversations that a stronger model classifier flagged as potentially harmful, and 94% of high-priority ones. The figures use a 1% false-positive rate relative to that classifier. They do not measure an independently established share of all genuinely harmful conversations.
Self-assessment was substantially weaker as an active blocker. In a controlled test of 24 cyber tasks with normal production classifiers disabled, attackers succeeded on 19 tasks without safeguards, 13 with self-assessment and six with a separate prompted classifier. Production self-assessment does not block conversations; it supplies signals for further review.
The researchers identify shared blind spots: the same teacher supplies training labels and the evaluation standard. A misaligned model could conceal its own actions in its assessment, while an honest model can repeat the same mistake. The team therefore treats the technique as an additional layer rather than a standalone defense.