Anthropic accused three major Chinese AI laboratories—DeepSeek, Moonshot AI, and MiniMax—of orchestrating “industrial-scale” campaigns to illicitly steal the capabilities of its Claude AI model.
The company alleges that these firms used a massive network of fraudulent accounts to bypass restrictions and harvest high-level reasoning and coding logic to train their own models.
The Details of the “Distillation Attack”
Anthropic claims the labs used a technique called “distillation,” where a less powerful model is trained on the high-quality outputs of a superior one (in this case, Claude).
- The Scale: Over 16 million exchanges were generated using approximately 24,000 fraudulent accounts.
- The Method: The labs allegedly used “proxy services” and “Hydra cluster architectures” (networks of fake accounts) to hide their identity and scale their queries across Anthropic’s API and third-party cloud platforms.
- Targeted Skills: The attacks specifically focused on Claude’s industry-leading features: complex reasoning, computer vision, agentic tool use, and high-end coding.
Specific Allegations by Lab
Anthropic broke down the volume and focus of each company’s alleged campaign:
| Company | Vol. of Exchanges | Primary Focus / Tactics |
| MiniMax | 13 Million+ | Focused on coding and tool use. Redirected 50% of traffic within 24 hours of a new Claude release to capture fresh data. |
| Moonshot AI | 3.4 Million | Targeted reasoning and computer vision. Identified via metadata matching public profiles of senior Moonshot staff. |
| DeepSeek | 150,000 | Focused on “Chain-of-Thought” data. Asked Claude to “articulate internal reasoning step-by-step” to train its own reasoning engines. |
National Security & Guardrails
The core of Anthropic’s concern is not just intellectual property, but the stripping of safety guardrails.
- Weaponization Risk: Anthropic warns that when Chinese firms “distill” American models, they inherit the raw intelligence but often remove the safety filters designed to prevent the AI from assisting in bioweapon development or cyberattacks.
- Export Control Circumvention: By siphoning intelligence via the cloud, these labs are effectively bypassing U.S. chip export bans meant to slow down the development of frontier Chinese AI.
Backlash: “The Glass House” Argument
The accusation has met with significant criticism and “hypocrisy” claims from the tech community, most notably from Elon Musk:
“Anthropic is guilty of stealing training data at a massive scale and has had to pay multi-billion-dollar settlements for their theft. This is just a fact.” — Elon Musk, Feb 23, 2026.
Critics point out that Anthropic itself is currently facing lawsuits for scraping data from the open internet (including Reddit and thousands of copyrighted books) to train Claude, leading some to argue that the company is now complaining about the very “unauthorized learning” tactics it used to build its own empire.
