Anthropic has Claude review its own conversations for misuse
The model-native monitor runs after responses across a large share of global traffic and adds less than 2% to inference compute.
Anthropic has deployed a technique called self-assessment that asks Claude to judge the safety of a conversation immediately after producing its normal response. The extra exchange is hidden from the user and feeds offline review rather than directly blocking an answer.
The approach reuses the conversation already held in the model KV cache. Anthropic estimates that each check generates about 10 additional tokens and adds less than 2% to inference compute, making broad coverage cheaper than sending every conversation to a separate classifier.
On Claude Fable 5, self-assessment flagged 78% of conversations identified as potentially harmful by a stronger prompted classifier, including 94% of the high-priority cases, at a 1% false-positive rate. Anthropic says it has used the system since June and now runs it in every Claude model released since Fable 5 across a large share of global traffic.
The company says the monitor has surfaced large-scale misuse that existing safeguards were not designed to detect. Because Claude uses its broader understanding rather than checking only a fixed list of harms, Anthropic expects detection quality to improve with the underlying model.
There are important limits. The assessment does not catch every harmful exchange, is not robust enough to stop a dedicated attacker, and does not replace real-time safeguards. It is a low-cost triage layer that can send suspicious conversations to stronger tools and human review.
