Claude defenses may have cut data harvested by distillers by around 90 to 99 percent
Anthropic says its measures limit access to Claude’s reasoning and use classifiers to block extraction attempts.
TLDR
Anthropic says early evidence suggests its new defenses have reduced the amount of data distillers harvest from Claude by around 90 to 99 percent. The measures include restrictions on where Claude’s reasoning can go, a two-stage classifier to block extraction attempts, and limits or bans for accounts used in repeated attempts. Anthropic says some developers have had to update their integrations, and its classifiers can wrongly block benign requests in fewer than 0.1% of cases.
Combined views
—
First seen ago