Anthropic Exposes Industrial-Scale AI Distillation Attacks by DeepSeek, Moonshot AI, and MiniMax
Original: Anthropic Exposes Industrial-Scale AI Model Distillation Attacks by DeepSeek, Moonshot AI, and MiniMax View original →
Industrial-Scale Distillation Attacks Discovered
On February 24, 2026, Anthropic publicly disclosed that major Chinese AI companies had been conducting large-scale distillation attacks against its Claude models. DeepSeek, Moonshot AI, and MiniMax were identified as the perpetrators.
Scale and Method
The attack involved:
- Creation of over 24,000 fraudulent accounts
- Generation of more than 16 million exchanges with Claude
- Using that conversation data to train and improve their own competing AI models
Why Illicit Distillation Is Dangerous
Anthropic distinguishes between legitimate and illicit distillation. While AI labs legitimately use distillation to create smaller, cheaper models for their customers, foreign labs that illicitly distill American models can remove safety guardrails and feed extracted capabilities into their military, intelligence, and surveillance systems.
Call for Coordinated Action
Anthropic warned that these attacks are growing in both intensity and sophistication, calling for rapid, coordinated action from industry players, policymakers, and the broader AI community to address the threat.
Full details are available in Anthropic's official report: Detecting and Preventing Distillation Attacks.
Related Articles
Anthropic is trying to make AI jailbreaks measurable, not just viral. Its July 2 framework separates minor bypasses from universal failures, adds a HackerOne path for Fable 5 reports, and says one new classifier blocks the Amazon-reported technique in over 99% of cases.
The Future of Life Institute’s Summer 2026 AI Safety Index grades nine frontier AI companies across 37 indicators, and no firm rises above C+. The sharper point is not who leads, but how weak the ceiling remains as model capabilities and defense use expand.
Anthropic says Claude contains a J-space that resembles a global workspace for active, verbalizable thoughts. The lead tweet has more than 9.1 million views and points to audit use cases, including hidden goals in sabotage-trained models.