Three Chinese AI Labs Ran Industrial-Scale Distillation Attacks on Claude, Extracting 16 Million+ Exchanges
Anthropic reveals that DeepSeek, Moonshot AI, and MiniMax used approximately 24,000 fraudulent accounts to conduct over 16 million exchanges with Claude, systematically extracting agentic reasoning, tool use, and coding capabilities for their own model training.
Anthropic disclosed on February 23 that three Chinese AI companies — DeepSeek, Moonshot AI, and MiniMax — conducted industrial-scale distillation campaigns against Claude. Using approximately 24,000 fraudulent accounts, the three firms collectively executed more than 16 million exchanges designed to extract Claude’s agentic reasoning, tool use, and coding capabilities for use in training their own models.
The Scale of Each Operation
MiniMax ran the largest campaign by far, accounting for over 13 million exchanges. The operation focused on extracting agentic coding and tool use capabilities, and was detected ahead of a new MiniMax model release. Anthropic observed a telling behavioral pattern: whenever a new Claude model was released, MiniMax redirected roughly half its traffic to the new model within 24 hours — a clear indicator of systematic capability harvesting.
Moonshot AI, the company behind the Kimi model series, conducted more than 3.4 million exchanges. Its operation targeted a broader range of capabilities: agentic reasoning, coding, computer-use agent development, and computer vision. In a particularly notable finding, Anthropic’s analysis of request metadata matched profiles of senior Moonshot AI staff, directly linking the company’s employees to the distillation infrastructure.
DeepSeek ran the smallest operation at over 150,000 exchanges, but its approach was methodologically distinct. Rather than simply extracting outputs, DeepSeek focused on reasoning capabilities and rubric-based scoring. It prompted Claude to “verbalize the internal reasoning of completed answers step by step” — effectively generating chain-of-thought training data at scale. DeepSeek also used Claude to generate censorship-safe alternative responses to politically sensitive queries.
How They Did It
All three companies used fraudulent accounts combined with commercial proxy services to access Claude at scale. Anthropic’s investigation found that a single proxy network was managing more than 20,000 fraudulent accounts simultaneously — suggesting a well-resourced, professionally operated infrastructure behind the distillation campaigns.
How Anthropic Caught Them
Detection relied on a combination of IP address correlation, request metadata analysis, infrastructure indicators, and intelligence from industry partners. The multi-signal approach allowed Anthropic to not only identify the distillation activity but also attribute it to specific organizations.
In response, Anthropic has deployed classifier systems and behavioral fingerprinting tools designed to detect and block distillation attempts in real time. The company indicated these systems are now running in production.
OpenAI Reports Similar Attacks
Anthropic isn’t alone. OpenAI has claimed that DeepSeek employed similar distillation techniques against ChatGPT, suggesting this is not an isolated incident but a systematic industry practice.
Model distillation — harvesting a model’s outputs at scale to train a competing model — is emerging as one of the most significant intellectual property challenges in the AI industry. Traditional software IP protections don’t map cleanly onto the problem: the “theft” happens through legitimate API calls, making detection and enforcement fundamentally different from conventional code piracy.
For AI providers, the implications are clear. Distillation defense is no longer optional — it’s a core infrastructure requirement. And as models become more capable, the incentive for competitors to extract that capability through distillation only grows.
Official announcement: https://www.anthropic.com/news/detecting-and-preventing-distillation-attacks
関連記事
中国AIラボ3社がClaudeに大規模蒸留攻撃、1600万回以上のやり取りで能力抽出を試行
Anthropicは、DeepSeek・Moonshot AI・MiniMaxの3社が約24,000の不正アカウントを通じて1600万回以上のやり取りを行い、Claudeの能力を組織的に蒸留していたことを公表した。
Anthropic、AIによるコード脆弱性検出ツール「Claude Code Security」を発表
AnthropicがClaude Code on the webに組み込まれた脆弱性検出機能「Claude Code Security」を限定リサーチプレビューとして公開。従来の静的解析を超え、人間のセキュリティ研究者のようにコードを推論し、複雑なビジネスロジックの脆弱性を検出する。
Claude Code v2.1.47 リリース - Windows対応強化・メモリ改善・40件超のバグ修正
Anthropicは2026年2月19日、Claude Code v2.1.47をリリースした。Windows環境の描画バグ修正、長時間セッションのメモリ改善、起動速度の向上(約500ms短縮)、コンテキスト圧縮後のプランモード保持など40件以上の修正が含まれる。
人気記事
Claude Code v2.1.93 リリース - PreToolUseフック延期決定やフリッカーFREE渲染など
AnthropicがClaude Code v2.1.93をリリース。PreToolUseフックの延期決定機能、フリッカー-freeレンダリングオプション、PermissionDeniedフック、名前付きサブエージェントのタイプアヘッド対応などを含む。
Claude Code v2.1.92 リリース - forceRemoteSettingsRefreshやBedrockセットアップウィザードなど
AnthropicがClaude Code v2.1.92をリリース。forceRemoteSettingsRefreshポリシー設定、Bedrockセットアップウィザード、costコマンドの改善など多数の機能追加とバグ修正を含む。
Claude Code v2.1.84 リリース - PowerShellツールプレビューと環境設定の強化
Claude Code v2.1.84がリリース。Windows向けのPowerShellツール、環境変数によるモデル選択のカスタマイズアイドルセッション処理の改善などが含まれる。
最新記事
Claude Code v2.1.93 リリース - PreToolUseフック延期決定やフリッカーFREE渲染など
AnthropicがClaude Code v2.1.93をリリース。PreToolUseフックの延期決定機能、フリッカー-freeレンダリングオプション、PermissionDeniedフック、名前付きサブエージェントのタイプアヘッド対応などを含む。
Claude Code v2.1.92 リリース - forceRemoteSettingsRefreshやBedrockセットアップウィザードなど
AnthropicがClaude Code v2.1.92をリリース。forceRemoteSettingsRefreshポリシー設定、Bedrockセットアップウィザード、costコマンドの改善など多数の機能追加とバグ修正を含む。
Claude Code v2.1.91 リリース - MCPツール結果の永続化とEditツールの最適化
Claude Code v2.1.91ではMCPツールの結果サイズの上書き、シェル実行の制御強化、Editツールの効率改善などが行われた。