Artificial Intelligence
Anthropic said it cut Claude reasoning theft by 99%
Anthropic said new controls reduced the extraction of Claude's internal reasoning tokens by 90 to 99 percent, while it linked large-scale distillation attacks to Zhipu and Moonshot and said it had banned tens of thousands of accounts.
- Anthropic said new defences reduced the extraction of Claude's internal reasoning tokens by 90 to 99 percent.
- The company said attackers had used networks of hundreds of thousands of fake accounts in distillation attacks.
- Anthropic named Zhipu and Moonshot as participants and said Zhipu had siphoned cyber capabilities for GLM-5.3 while Moonshot had routed queries through Claude.
- Anthropic said it now limited access to Claude's internal reasoning steps, used a classifier to spot extraction attempts, and automatically banned flagged accounts.
- It said tens of thousands of accounts had already been banned, and that legitimate users were also being caught.
Anthropic said it was facing large-scale distillation attacks on its Claude models. It said those attacks copied a model's capabilities to train smaller systems, and that its new controls reduced extracted reasoning tokens by 90 to 99 percent.
The company said attackers used networks of hundreds of thousands of fake accounts. It named Zhipu and Moonshot as participants in the activity.
Anthropic said Zhipu had siphoned cyber capabilities from leading US models for its GLM-5.3 model. It said Moonshot had routed queries through Claude without disclosure.
Anthropic said the stolen capabilities could be used in foreign military and surveillance systems. It also said it had limited access to Claude's reasoning steps, added a classifier for extraction attempts, and banned flagged accounts automatically.
The company said tens of thousands of accounts had already been banned. It added that some legitimate users were being caught up in the process too.
Named in this story
Companies
- Anthropic
- said it cut extraction of Claude reasoning tokens and banned accounts
- Zhipu
- was named by Anthropic as a participant
- Moonshot
- was named by Anthropic as a participant
Products and systems
- Claude
- the model family targeted by the distillation attacks
- GLM-5.3
- the model Anthropic said had received siphoned capabilities
How the source tells it
The source mixed technical explanation with alarmist and promotional framing, using scale language and Anthropic’s own claims to raise the stakes.
- alarm loaded scale language and a worst-case military use claim
- hype the company’s 90 to 99 percent reduction claim was repeated without outside verification