Prologue: The Most Diligent Student

One day in February 2026, Anthropic's security team detected an anomalous signal. Somewhere, an extraordinarily systematic flood of questions was being directed at Claude — not out of curiosity, but with clear intent. "Reason through this problem step by step." "Complete this coding task using an agentic approach." "Explain exactly how to use this tool." Every query was precisely calibrated to extract Claude's most differentiated capabilities.

The investigation revealed that three Chinese AI companies — DeepSeek, Moonshot AI, and MiniMax — had deployed over 24,000 fraudulent accounts to conduct more than 16 million exchanges with Claude. This article covers the how: the methods, the infrastructure, and the full anatomy of the attack.

1. What Happened: The Overview

Timing and Context

Anthropic published its findings on February 23, 2026. The timing was notable. The disclosure came as the US government was actively debating AI chip export controls, and OpenAI simultaneously submitted a letter to Congress warning of "industrial-scale distillation campaigns" by Chinese AI firms.

As of the time of reporting, DeepSeek, Moonshot AI, and MiniMax had not issued official responses.

The Numbers

Company	Exchanges	Primary Target Capabilities
DeepSeek	150,000+	Logical reasoning, reward modeling (rubric grading), censorship-safe query generation
Moonshot AI	3.4M+	Agentic reasoning, tool use, coding, computer vision
MiniMax	13M+	Agentic coding and tool use (largest volume)
Total	16M+	24,000+ fake accounts

2. The Core Method: What Is Model Distillation?

Where Legal Ends and Illegal Begins

Model distillation is a legitimate machine learning technique: train a smaller, less capable model on the outputs of a larger, more capable one. When applied to your own models, it is entirely lawful and widely used across the industry to produce efficient, smaller versions of larger systems.

The line is crossed when a competitor's model outputs are used without authorization. Claude's API Terms of Service explicitly prohibit using API responses to train competing models. This case represents what may be the largest documented, systematic violation of that boundary to date.

DeepSeek's Distinctive Technique: Make It Think Out Loud

DeepSeek appears to have used a method distinct from the other two companies: Chain-of-Thought Elicitation. By prompting Claude to reason through problems step by step — "think through this carefully before answering" — DeepSeek harvested not just answers but Claude's reasoning process itself as training data.

Anthropic also noted that DeepSeek's queries included requests for help generating "censorship-safe alternatives" to politically sensitive topics — questions about dissidents, party leadership, and authoritarianism — with the apparent goal of bypassing content restrictions in their own model.

3. How They Tried to Stay Hidden: The Hydra Cluster

Infrastructure Design

All three campaigns shared a similar infrastructure pattern that Anthropic termed the "Hydra Cluster" architecture. The core structure:

[Attacker Server]
       │
       ▼
[Proxy Network] ──── manages 20,000+ fake accounts simultaneously
    │         │         │
    ▼         ▼         ▼
[Claude API] [Claude API] [Claude API]
 (Account A)  (Account B)  (Account C)

A single network managed over 20,000 fake accounts concurrently, distributing traffic so that each individual account looked within normal usage bounds. If one account triggered a detection signal, the rest of the cluster continued operating uninterrupted — a structure named after the mythological Hydra, whose severed heads always grew back.

Evasion Tactics

Traffic distribution: Kept per-account query volumes within normal user ranges
Proxy routing: Concealed real IP addresses behind proxy services
Credential cover: Accounts registered under the guise of legitimate research or educational use

4. How Anthropic Detected Them

Detection Methodology

Anthropic disclosed four signal types used in cross-analysis to identify the campaigns:

IP address correlation: Different accounts accessing Claude from the same IP ranges or proxy services
Request metadata analysis: Abnormal regularity in headers, timestamps, and request patterns
Behavioral fingerprinting: Identification of repetitive, systematic query patterns across accounts
Synchronized traffic: Similar queries fired simultaneously from coordinated account clusters

Anthropic also noted that intelligence sharing with industry partners contributed to detection — suggesting that information from OpenAI and other firms helped triangulate the attack patterns.

Response Actions

Built and deployed classifiers for distillation attack pattern detection
Strengthened verification requirements for educational and research account registrations
Developing model-level safeguards to reduce distillation signal quality
Shared technical indicators with industry partners and authorities

5. Six-Month Outlook: Three Scenarios

Scenario 1: Regulatory and Legal Escalation (Probability: 65%)

Anthropic's public disclosure is likely to be used as evidence in US Congressional debate over AI chip export controls. OpenAI's concurrent letter to Congress reinforces this direction. However, actual litigation faces significant hurdles: jurisdiction over Chinese entities is complex, and the legal protectability of AI model outputs remains unsettled.

Scenario 2: Arms Race in API Defense Technology (Probability: 80%)

With distillation attacks now officially confirmed at scale, major AI companies are expected to accelerate investment in API-level defenses — output watermarking, response pattern perturbation, and real-time anomaly detection. These measures may affect API service complexity and cost for legitimate users.

Scenario 3: Chinese AI Labs Pivot to Independent Data Strategies (Probability: 55%)

In the short term, evasion techniques may be refined. Over the medium term, if access to US AI models becomes significantly restricted, there is likely growing pressure to develop independent synthetic data generation methodologies — reducing reliance on frontier model distillation altogether.

6. Decision-Making Guide

If You Provide an AI API Service

Check Question	If Yes: Priority Action
Does your ToS prohibit use of API responses for competing model training?	Establish a ToS violation monitoring framework
Do you have anomalous bulk API call detection in place?	Implement per-account query pattern anomaly detection
Have you evaluated output watermarking technology?	Begin assessment of model output traceability options
Do you share anomalous traffic intelligence with industry peers?	Explore joining AI security information sharing channels

If Your Organization Uses the Claude API

Check Question	If Yes: Priority Action
Have you reviewed Claude's API Terms of Service recently?	Check for material changes to usage restrictions
Is your API key management policy documented?	Audit internal API key access permissions
Do you know the training data provenance of external AI tools you use?	Conduct an AI supply chain risk review

7. What Not to Overestimate

Risk 1: "Distillation made them equivalent to Claude"

Training on 16 million queries does not mean Claude's full capability set was replicated. Distillation can meaningfully improve specific skills, but it does not transfer overall alignment quality, the breadth of a knowledge base, or general reasoning depth. Anthropic itself flagged that the distilled models likely absorbed targeted capabilities without the corresponding safety guardrails — a concern, but a different one than wholesale capability transfer.

Risk 2: Generalizing to all Chinese AI companies

The accusations are explicitly limited to three companies: DeepSeek, Moonshot AI, and MiniMax. Extending this to characterize the entire Chinese AI ecosystem would be an unsupported generalization.

Risk 3: Casting Anthropic as a straightforward victim

Several outlets, including Futurism, noted that Anthropic itself faces questions about the transparency of data used to train Claude. This is part of a broader structural debate about training data ethics across the AI industry — one that applies well beyond this specific incident.

8. Epilogue: The First Official Salvo in a War Without Rules

Model distillation from competitor systems has long been an open secret in the AI industry — a gray zone everyone knew about but no one named publicly. Anthropic's disclosure represents the first attempt to formally draw a line in that gray zone.

Whether that line will hold remains unclear. Whether it carries legal force, whether it leads to regulation, or whether it simply triggers a technical arms race — the shape of those outcomes will only emerge over the coming months.

One thing seems certain: protecting AI model capabilities has moved beyond the scope of API terms of service and into the domain of geopolitics and industrial policy. Part 2 of this series examines the structural conditions that made this possible — the architecture of the US–China AI competition and why this gray zone existed in the first place.

Key Action Summary

Role	Check Now	Review Within 3 Months
AI Service Developer	Confirm ToS prohibition on competitor model training	Evaluate anomalous traffic detection logic
Enterprise AI Adoption Lead	Review terms of external AI services currently in use	Build AI supply chain risk assessment framework
AI Policy / Strategy Lead	Read Anthropic's primary disclosure and OpenAI's Congressional letter	Monitor US AI chip export control developments
Security Lead	Audit API key management and access permissions	Explore AI security information sharing channels

Frequently Asked Questions

Q1. Is model distillation illegal by default?

No. Distillation is a legitimate machine learning technique. Training a smaller in-house model on outputs from a larger in-house model is standard practice. What makes this case problematic is the unauthorized use of a third party's (Anthropic's) model outputs. Claude's API Terms of Service explicitly prohibit using API responses to train competing AI models.

According to Anthropic's disclosure, MiniMax focused on agentic coding and tool use. Coding-related tasks tend to require multiple iterative exchanges to produce a single usable output, naturally inflating query counts. That said, a higher query count does not proportionally equal greater capability transfer.

Q3. Can Anthropic pursue legal action?

Potentially, but with significant complications. The most viable path would be a civil lawsuit for ToS breach and related damages. However, the defendants are Chinese entities, raising complex jurisdictional issues. The legal protectability of AI model outputs under copyright law also remains unsettled. It is also plausible that the public disclosure was designed to generate regulatory and political pressure rather than serve as a prelude to litigation.

Q4. Can these attacks be fully prevented?

The industry consensus is that complete prevention is technically difficult. Anthropic has committed to strengthening detection and defense, but sophisticated attackers can adapt their methods as defenses evolve. Watermarking, response perturbation, and real-time anomaly detection are all active research areas — but no comprehensive solution exists today.

Series Guide

Part 1 (this article): Methods & Tech — How it was done

Part 2 (scheduled 2026-02-26): Structure & Competition — Why it was possible (the gray zone of US–China AI rivalry)

Part 3 (scheduled 2026-02-27): Regulation & Future — How AI model protection will change

Update Notes

Content reference date: 2026-02-25 (KST)
Update cadence: as significant developments occur
Next scheduled review: 2026-03-10

References

Anthropic official disclosure: https://www.anthropic.com/news/detecting-and-preventing-distillation-attacks
TechCrunch: https://techcrunch.com/2026/02/23/anthropic-accuses-chinese-ai-labs-of-mining-claude-as-us-debates-ai-chip-exports/
CNBC: https://www.cnbc.com/2026/02/24/anthropic-openai-china-firms-distillation-deepseek.html
Fortune: https://fortune.com/2026/02/24/anthropic-china-deepseek-theft-claude-distillation-copyright-national-security/
The Hacker News: https://thehackernews.com/2026/02/anthropic-says-chinese-ai-firms-used-16.html

claude-theft Posts

16 Million Queries: How China's AI Labs Used Claude as a Textbook