Threat Intelligence

Anthropic Uncovers Large-Scale Effort to Clone Its Claude AI Model

The Hacker News · 12 Sept 2026

Key Takeaway Businesses using AI tools should review vendor terms carefully and be cautious about which platforms process sensitive data, since conversations may be captured and reused without your knowledge.

Anthropic has disclosed that it identified and disrupted large-scale attempts by seven China-based AI labs to illicitly copy the capabilities of its Claude AI model. The labs named include Alibaba, Moonshot, DeepSeek, Z.ai (also known as Zhipu), and MiniMax.

The technique involved, known as knowledge distillation, is normally a legitimate way to train smaller AI models using outputs from larger ones. However, Anthropic says these labs used illicit versions of the process, covertly extracting Claude's responses without authorisation and using them to train their own competing models. This was done through networks of fake accounts created with stolen credit cards, login credentials, and API keys, often routed through proxy services that disguise the true source of the requests.

Anthropic said some labs fed real conversations between their own users and their models into Claude, then used Claude's responses as training data, in some cases exposing sensitive information from individuals, major companies, and state-affiliated actors. In other instances, user requests were secretly rerouted to Claude without the users' knowledge or consent, purely to harvest exchanges for training purposes.

AI security data privacy Anthropic AI model theft third-party risk

Summarised by CISO AI from The Hacker News. We link back to every original so you can read it yourself.