Anthropic deploys Claude self-assessment safety monitoring
Anthropic says its self-assessment technique adds about 10 tokens per answer and has flagged 94% of high-risk conversations in internal tests.
Anthropic has disclosed a safety monitoring technique called self-assessment, in which Claude reviews its own conversation immediately after answering, according to aitimes.com. The company said the method has been applied to a significant share of Claude usage worldwide since Claude Fable 5 launched in June.
Anthropic said self-assessment adds an average of about 10 tokens per exchange, keeping the extra compute cost under 2% of total inference cost. In an internal evaluation using real service conversations, Fable 5 flagged 78% of conversations that an existing classifier had judged harmful.
The detection rate rose to 94% for high-risk conversations, with a 1% false-positive rate measured against that classifier. Anthropic said the technique also works on risk types excluded from training, such as fraud and cyberattacks.
Earlier detections cited by the company
Anthropic said that in August it detected more than 100,000 suspected child sexual exploitation activities on a single API customer account and reported them to authorities. In September, the company said it identified and blocked a group that used Claude Code to automate romance scams against dozens of people.
On October 29, 2026, Anthropic posted on X that it had trained Claude to flag potential misuse through self-assessment.
Quick answers
What is Anthropic's self-assessment technique?
It is a safety monitoring method in which Claude reviews its own conversation right after answering to flag potential misuse.
How much does self-assessment add to inference cost?
Anthropic says it adds an average of about 10 tokens, keeping extra compute cost under 2% of total inference cost.
When did Anthropic start using self-assessment?
The technique has been in use since Claude Fable 5 launched in June.