Anthropic's Claude adds self-assessment feature scanning conversations since June 2026
Technology
Anthropic's AI, Claude, just got a new "self-assessment" feature that quietly scans conversations for misuse or harmful content.
Running since June 2026, it spots tricky problems older safeguards missed and keeps getting smarter as the AI learns.
Anthropic's Claude flags only for review
This upgrade is way more efficient, using less than 2% extra computing cost (down from nearly 24%) and catches about 78% of harmful chats and almost all top-priority cases.
In August alone, it flagged over 100,000 instances of child exploitation material for one customer.
Still, it only flags issues for review instead of blocking them instantly, and hackers can sometimes slip by.
That's why Anthropic uses this tool alongside stronger real-time protections to keep things safer overall.