Anthropic Strengthens Defenses Against Distillation Attacks on Claude

At a glance

Anthropic said it has introduced new defenses to stop 'distillation attacks,' in which other companies secretly copy Claude's abilities. Early data suggests the method cut the data taken by distillers by about 90 to 99 percent. The change also caused inconvenience for regular users, who must now change how they connect to the service.

Key points

1. 1

   A distillation attack (a method of taking another AI's abilities without permission to build a similar model) is often carried out on a large scale using many fake accounts.

2. 2

   Models created through distillation may lack sufficient safety guardrails, which could lead to harm in biology or cybersecurity, experts warn.

3. 3

   The article says Chinese company Zhipu tried to distill Claude, and that the resulting GLM-5.3 model's safety features could be easily bypassed.

4. 4

   The article also cites a case where distillation could become a national security issue. A user who appeared to be linked to the Chinese People's Liberation Army reportedly asked for CCTV surveillance analysis of a…

Why it matters

Secretly copying AI models is not just a problem for one company. It is directly tied to AI safety and national security, which is why it matters.