Anthropic says Claude now screens its own conversations for misuse, matching 78% of a stronger classifier's flags

Anthropic · Research report

Read the source (opens in a new tab)
What happened
Anthropic described self-assessment on its Safeguards blog on October 16, 2026: after each response, Claude writes a short safety assessment of the conversation, which feeds offline misuse review.
Key points
  • Anthropic runs it in every Claude model released since Fable 5, across a large fraction of global traffic, and estimates it adds less than 2% to inference compute.
  • On Fable 5 it flags 78% of conversations a stronger prompted classifier flags as potentially harmful, and 94% of high-priority ones, at a 1% false-positive rate.
  • In August it raised over 100k child sexual abuse and exploitation flags from one API customer, reported to NCMEC; it also exposed a romance scam built with Claude Code.
  • Flags outside Anthropic's harm categories included illicit attempts to extract Claude's reasoning, prompt injection, and agents taking actions their users had not authorized.
Evidence
A research post by nine Anthropic authors on its own production safeguard. Recall is measured against Anthropic's own prompted classifier, not confirmed harm; the enforcement cases are the company's account.
Uncertain
Anthropic says it "does not catch every harmful conversation" and is not robust to an attacker targeting it directly; a model hiding a misaligned action "could hide it in the assessment too."

See it in the latest record