The company estimates that, after implementing the new safeguards, reasoning tokens extracted by distillers on its most capable generally available models fell by a factor of 10 to 100 or more; on some days, the amount was too small to measure. Anthropic says it has banned tens of thousands of accounts associated with distillation. It also reports a classifier false-positive rate below 0.1% and says blocked requests in specified areas, including distillation, have been billable since September 24. The company cautions that attackers may find ways around the defenses and acknowledges trade-offs for legitimate users. These figures and assessments are Anthropic’s own reports about its platform.
Strengthening our defenses against distillation attacks \ Anthropic
In one minute
Anthropic said it has expanded defenses against attempts to extract Claude’s reasoning for model distillation, combining controls on reasoning context, a two-stage classifier, and stricter limits and bans for persistent attempts.
What changed: Same event as Strengthening our defenses against distillation attacks \ Anthropic