Anthropic described a self-assessment step in which Claude adds a compact safety judgment after replying, averaging around ten tokens and an estimated compute overhead below 2%. On Fable 5, it flagged 78% of potentially harmful conversations and 94% of high-priority ones identified by a stronger prompted classifier, at a 1% false-positive rate. Those are not detection rates for all human-confirmed misuse. Used operationally since June, the offline monitor helps prioritize investigation but does not replace real-time safeguards or independent oversight and remains vulnerable to determined evasion.