Anthropic称多家中国AI实验室盗用超亿条Claude对话训练模型

This guide covers the key facts of the recent high-profile AI intellectual property dispute, plus practical tips to protect your personal information when using AI products.

1. Core event overview: On September 11, 2026, U.S. AI firm Anthropic released a threat intelligence report disclosing that it had blocked multiple waves of unauthorized access from Chinese AI labs between December 2025 and August 2026. The unlicensed use of Claude model outputs to train competing in-house models, defined as "illicit distillation," involved more than 100 million interactions. The report also identified six additional categories of blocked high-risk activity, including cyber operations, fraud, and biological misuse, and named leading Chinese AI players including Alibaba, Moonshot, DeepSeek, and Xiaomi.

2. Specific conduct of named firms: An Alibaba-affiliated entity made more than 151 million Claude API calls between May and July 2026 to train its Qwen model, at one point using over 3,500 fraudulent accounts to generate nearly 3 million requests per day at peak volume. Moonshot, developer of the Kimi chatbot, secretly routed user queries to Claude and returned results to users without their knowledge, logging more than 23 million calls over three months and retaining content to train its own model. DeepSeek used the same tactic, generating over 12 million requests in just 14 days in July 2026.

3. Practical user guidance: Content routed by the named firms included large volumes of personal and business sensitive information. When using the affected Chinese AI products, avoid submitting private or confidential material. As of the report’s release, none of the implicated companies have issued public responses; users may follow developments as the situation evolves.

The recent illicit model distillation allegations offer clear takeaways for AI-focused brands across compliance, user trust, and product development.

1. Compliance risk warning: Unauthorized use of third-party large language model (LLM) outputs to train in-house models constitutes clear IP infringement, and may also violate privacy regulations and contradict a company’s own terms of service. Several implicated firms hid from users that their queries were being routed to third-party models, violating user知情权 (right to know) and exposing sensitive information across multiple stakeholders—risks that directly erode a brand’s reputation for independent R&D and public trust. Per disclosed figures, an Alibaba-affiliated party made over 151 million Claude calls across three months, using outputs for core R&D workflows including reinforcement learning and model architecture design. If confirmed, such conduct would cause severe reputational damage.

2. User sentiment insight: The incident highlights that AI users are highly sensitive to how their content is handled and who actually provides the services they use. Brands must clearly disclose service providers and data use purposes in marketing and service delivery; deliberate concealment carries a high risk of triggering a trust crisis.

3. R&D boundary guidance: While stealing outputs from leading models to close technical gaps can reduce costs and accelerate development in the short term, it carries extreme compliance risk. Brands must balance R&D efficiency with clear compliance guardrails to avoid similar IP disputes.

The recent AI IP infringement incident provides actionable risk warnings, opportunity identification, and response guidance for sellers operating in AI-related segments.

1. Clear operational risks: Sellers of AI products and services that access third-party LLMs without authorization or secretly route user queries to train their own products face access blocks and IP infringement lawsuits from model owners. Firms in this incident used large volumes of fraudulent accounts registered with addresses in Singapore and Japan to bypass restrictions. Such tactics may not only violate privacy regulations but also trigger user claims over violated知情权 (right to know), with additional liability for sensitive data leaks.

2. Market opportunity identification: The scale of the implicated activity signals nine-figure demand for high-performance, low-cost LLM access. DeepSeek previously gained rapid market traction on the strength of its strong performance and low pricing, pointing to clear growth potential for licensed LLM APIs and compliant training data services. AI products positioned around zero data leakage and transparent service chains will also see stronger user preference.

3. Incident response priorities: Sellers of AI tools should audit their model supply chains immediately, clearly disclose data flows to users, and secure official model licenses in advance to mitigate infringement risk.

The allegations that Chinese AI firms engaged in illicit distillation of overseas LLMs offer strategic guidance and risk warnings for manufacturers developing AI-enabled products and pursuing digital transformation.

1. Product design and manufacturing guidance: Disclosed industry details show extremely strong demand among Chinese AI firms for high-performance, low-cost LLM capabilities, to the point that some firms resorted to rule-breaking tactics to access leading model outputs. When developing smart hardware or embedded AI features, manufacturers should prioritize licensed, compliant LLMs as their technical base to avoid IP disputes affecting product launches and brand reputation.

2. Digital transformation risk warning: When adopting AI tools to optimize production or build digital systems, manufacturers must rigorously vet AI service providers for compliance. In this incident, some providers secretly routed user queries to third-party models, creating risks of leakage for sensitive information such as production parameters and commercial secrets. Contracts with AI vendors should clearly define data flow terms and confidentiality obligations.

3. Commercial opportunity reference: Clear market gaps exist for hardware supporting compliant model training, data security protection hardware, and equipment for localized AI deployment. Manufacturers may target these segments based on their existing production capacity.

Anthropic’s illicit distillation threat intelligence report clearly outlines existing pain points in the AI industry, pointing to clear business expansion directions for AI service providers.

1. Industry trend assessment: The AI sector remains in a period of rapid iteration, with widespread demand among firms to catch up to leading model capabilities and reduce training costs. However, unauthorized distillation has been explicitly classified as IP infringement by top model vendors, who are actively blocking such activity via technical measures. Industry compliance is a clear long-term trend; the gray-area practice of scraping third-party model outputs to train in-house models is no longer viable.

2. Client pain point mapping: Per disclosed volumes, the three named firms alone—Alibaba, Moonshot, and DeepSeek—generated over 186 million distillation-related interactions, pointing to three core pain points for AI enterprises: high costs for high-performance LLM access, insufficient supply of compliant training data, and gaps in anomalous access risk control. B2B clients also face risks of data being routed and leaked across the AI service chain; content forwarded in this incident included large volumes of sensitive information from multinational corporations.

3. Solution direction: Providers can develop targeted offerings across three categories to match market demand: official LLM license integration, compliant training data supply, and end-to-end AI chain data encryption and traceability services.

The large-scale illicit distillation incident disclosed by Anthropic offers actionable guidance for platform operators to refine risk control rules and mitigate operational risks.

1. Risk control optimization priorities: To bypass access restrictions, implicated firms registered large volumes of fraudulent accounts with listed addresses in Singapore and Japan, generating nearly 3 million anomalous interactions per day across more than 3,500 accounts at peak volume. This demonstrates that bad actors use bulk registration of geofaked accounts to evade controls. Platforms should strengthen account registration verification and anomalous traffic detection rules, referencing Anthropic’s risk detection mechanisms to block bulk abnormal requests in real time.

2. Operational management requirements: When providing AI services or integrating third-party AI capabilities, platforms must fulfill user disclosure obligations. In this incident, Moonshot and DeepSeek routed user queries without notice, not only violating user知情权 (right to know) but also creating sensitive information leakage risks. When integrating third-party capabilities, platforms should provide clear disclosures on service pages and in user agreements, and add sensitive information filtering mechanisms.

3. Risk mitigation guidance: When onboarding AI service providers, platforms must rigorously audit their model IP compliance to avoid associated liability from infringing vendors. Platforms training their own AI models must also avoid unauthorized use of third-party outputs to guard against infringement claims.

Anthropic’s illicit distillation report targeting Chinese AI labs provides a salient new case study and research directions for AI industry and governance research.

1. Emerging industry trend observation: The global LLM industry has entered a technology diffusion phase, where technical barriers between leading firms and challengers are being eroded via model distillation. Leading Chinese AI players including Alibaba, Moonshot, DeepSeek, and Xiaomi have not only used distillation to acquire leading model capabilities for model training, but also applied distillation outputs to core R&D processes including reinforcement learning and model architecture design. Some firms directly route user queries to leading models and pass off results as their own proprietary service; DeepSeek’s rapid market share gains, driven by its strong performance and low cost, were substantially supported by such tactics.

2. Emerging industry issue mapping: Unauthorized distillation creates three new categories of risk. First, LLM intellectual property boundaries remain ambiguous, with no clear standard for determining infringement when model outputs are used to train competing models. Second, data security risks are acute: routed content includes sensitive personal, corporate, and stakeholder information, creating significant privacy compliance exposure. Third, unfair competition risks exist: firms that gain cost and performance advantages via illicit distillation crowd out market space for companies that invest in compliant R&D.

3. Policy implications: Regulators should accelerate clarification of rules governing LLM IP protection, AI user知情权 (right to know) guarantees, and training data compliance to guide standardized industry development. The fact that no implicated firm had issued a public response as of the report’s release also reflects gaps in existing cross-border AI IP dispute resolution mechanisms.