Anthropic Cuts System Distillation Reduction by up to 99%
The Facts
- Anthropic said Monday that new measures to keep rivals from copying its Claude AI systems have reduced data harvested in so-called distillation attacks by about 90% to 99%, a concern heightened by U.S.-China competition in AI.
- The company said the volume of reasoning tokens obtained by suspected distillers is now 10 to 100 times lower than before, and that it has terminated tens of thousands of accounts associated with suspected distillation activity.
- A feature called "preserved thinking," introduced with Claude Fable 5.1 and Opus 5.5, prevents API accounts created on or after Aug. 31, from editing Claude's prior context, a technique Anthropic alleged attackers used to extract reasoning transcripts.
Sources Split
LEFT
CENTER
RIGHT
+2
+1
+1
ESTAB
CRIT
CENTER
PRO
ESTAB
The Spin
Pro-industry narrative
Blocking industrial-scale distillation prohibits knockoff AI systems from stripping the safeguards that keep frontier capabilities away from bioweapon and cyberattack work. Letting authoritarian militaries and surveillance agencies piggyback on American research is a national security failure that can't be allowed to continue, with Anthropic's measures a key step towards keeping the entire world safe.
Industry-critical narrative
Describing distillation as theft rings hollow coming from labs that hoovered up books, journalism and video without asking anyone first. Distillation is routine training practiced openly across Silicon Valley, and the alarm looks like a competitive moat wearing a national security badge. Rules that apply to one side of the Pacific and not the other are not rules at all.
