Anthropic reports on October 12 that new safeguards have reduced the Claude reasoning extracted by attackers on its own platform by roughly 90 to 99 percent.
The measures bind reasoning to its original context, detect extraction attempts with classifiers and restrict repeated attempts through usage limits and account bans.
The company acknowledges costs: developers must update integrations, legitimate requests can be wrongly blocked and some blocked requests are charged.
These are Anthropic's results. It does not yet know how long the gains will last and has not observed a highly sophisticated, persistent attempt to bypass the new safeguards.