Artificial intelligence / news brief
Anthropic uses Claude to flag risky conversations after each response
- Event
- Published

Anthropic says Claude now produces a short safety assessment after each response across a large share of global traffic. On Fable 5, the system flagged 78% of conversations caught by a stronger classifier and 94% of high-priority cases at a 1% false-positive rate, using under 2% extra inference compute.
Why it matters
Reusing the serving model could widen misuse monitoring cheaply, but Anthropic says the system is an offline triage layer and is not robust enough to block dedicated attackers.
Sources
- Scaling Misuse Detection with Self-Assessment (external site)Anthropic · primary source ·
- Claude Self-Assessment: Anthropic's Cheap Misuse Monitor (external site)CellCog · coverage ·
- Anthropic explains how Claude flags conversations that may involve harmful activity (external site)Business Today · coverage ·
- Inside the spy plot around Nvidia's CEO (external site)The Rundown AI · coverage ·