Who's Really Stealing From Whom?

2026-04-17 · Source: What's AI by Louis-François Bouchard · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation, Robotics & Autonomous Systems · Depth: Intermediate, long

Summary

Anthropic, OpenAI, and Google recently accused Chinese AI labs Deepseek, Moonshot AI, and Minimax of "industrial scale distillation attacks" on their frontier models, specifically Claude and Gemini, between February 12-23, 2026. Anthropic's detailed report described "Hydra clusters" using 24,000 fake accounts and 16 million exchanges to extract capabilities like reasoning, coding, and agentic behavior. Deepseek allegedly focused on reasoning extraction, rubric-based grading for reinforcement learning, and censorship-safe alignment. Minimax primarily targeted agentic coding and tool use, while Moonshot AI focused on agentic reasoning and computer vision. This blackbox distillation involves querying models at scale via APIs to harvest outputs for training, a technique formalized in 2015 by Hinton, Vignels, and Dean.

Key takeaway

For CTOs and VPs of Engineering evaluating model development strategies, recognize that frontier model intelligence is increasingly commoditized and extractable via public APIs. While distillation offers cost-effective model training, be aware of the unclear legal landscape regarding terms of service violations versus IP law. Focus on building value into product ecosystems and workflows around models, rather than solely on raw model outputs, to mitigate the impact of widespread distillation.

Key insights

Blackbox distillation allows smaller models to learn from frontier models' API outputs, even without access to internal probabilities.

Principles

• Model intelligence is extractable via public APIs.
• Scaling blackbox distillation yields significant performance gains.

Method

Query a teacher model with millions of carefully chosen questions, collect its text outputs (answers, explanations, code), then train a student model to imitate these patterns.

In practice

• Use frontier models as reward models for RLHF.
• Generate synthetic reasoning traces for training thinking models.

Topics

• Model Distillation
• Blackbox Distillation
• Frontier Models
• AI Ethics
• Copyright Infringement

Best for: CTO, VP of Engineering/Data, Research Scientist, AI Scientist, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• AI Distillation Explained: The Truth Behind the Biggest AI Controversy Right Now — AI model distillation attacks are a significant and escalating threat to intellectual property and frontier model capabilities.
• Owning the AI Pareto Frontier — Jeff Dean — Balancing frontier and efficient AI models through distillation and energy-aware hardware co-design is crucial for scaling AI capabilities.
• Mustafa Suleyman's case against open-source AI shortcuts — Distillation limits open-source AI's general utility, creating a significant performance gap with frontier models.
• Elon Musk testifies that xAI trained Grok on OpenAI models — AI model distillation, a method of training new models from existing APIs, is a widespread industry practice.
• Model Distillation Guide: Compressing LLMs for Edge Efficiency — Model distillation compresses large language models into smaller, efficient versions using a teacher-student framework.
• How much does distillation really matter for Chinese LLMs? — Distillation, while a legitimate training method, is being used illicitly by Chinese labs to extract capabilities from US frontier AI models.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by What's AI by Louis-François Bouchard.