Anthropic says defenses cut Claude data extraction by 90 to 99 percent

Anthropic reported on October 12 that defenses against unauthorized model copying reduced data extracted from Claude by approximately 90 to 99 percent, based on internal platform metrics. These campaigns query models at scale to reproduce their capabilities in another model. The defenses restrict access to reasoning transcripts, classify extraction attempts and limit or ban abusive accounts. Anthropic says it has banned tens of thousands of accounts associated with these campaigns. Some developers have had to update integrations, and benign requests can be blocked in fewer than 0.1% of cases. Users are charged for those blocked requests. Anthropic says the defenses have not yet faced a highly sophisticated, persistent circumvention attempt.