Anthropic Alleges Massive Data Scraping by Alibaba-Linked Accounts

AI startup Anthropic has requested that the U.S. Congress investigate a series of large-scale attempts to scrape its proprietary Claude AI models. In a recent letter to American senators, the company alleged that actors associated with the Chinese e-commerce giant Alibaba used nearly 25,000 fraudulent accounts to interact with Claude approximately 29 million times between late April and early June 2024. Anthropic claims these interactions were designed to extract the internal logic and architecture of its software, a process known in the industry as model distillation.

What is model distillation and why does it matter?

Model distillation is a technique where developers train a smaller, more efficient AI model using the output of a larger, more powerful one. By repeatedly querying a sophisticated model like Claude and analyzing its responses, a third party can effectively “clone” the performance characteristics of the original software without access to the underlying training data or source code. According to the Financial Times, which first reported on the letter, Anthropic contends that these unauthorized activities violate its terms of service and threaten its competitive edge in the global AI market.

How is this impacting U.S.-China technology policy?

The allegations have intensified calls for stricter export controls on advanced technology. In its communication with Congress, Anthropic urged lawmakers to restrict the access of Chinese entities to high-end AI chips and requested that the government grant American AI firms legal immunity to collaborate on identifying and blocking such scraping operations. This position places Anthropic at odds with hardware manufacturers like Nvidia, which has historically lobbied the U.S. government to maintain access to the Chinese market for its advanced processor exports.

How does this compare to previous AI industry controversies?

The practice of using one AI model to train another has become a contentious issue among major tech players. While Anthropic characterizes the alleged scraping as a security threat, the industry has seen similar behavior acknowledged in legal proceedings. For instance, Elon Musk admitted during a 2024 court deposition that his company, xAI, utilized outputs from OpenAI’s models to assist in the training of his own chatbot, Grok. While the circumstances differ—Musk’s admission involved an internal development strategy, whereas Anthropic’s allegations describe an external, potentially illicit campaign—the events highlight the growing difficulty in protecting intellectual property within the rapidly evolving generative AI landscape.

What is the status of Anthropic’s own regulatory challenges?

Anthropic is currently managing a separate, high-profile dispute with the U.S. government regarding its own AI safety protocols. Federal regulators recently paused the deployment of a specific Anthropic model, internally referred to as “Fable 5,” citing concerns that its safety guardrails could be bypassed. Specifically, the government expressed apprehension that the model’s capability to analyze software for vulnerabilities might be repurposed for malicious cyberattacks. Anthropic has publicly stated that the identified risks are manageable and continues to engage in discussions with officials to resolve the temporary suspension.

Key Takeaways

• The Allegation: Anthropic claims that 25,000 accounts linked to Alibaba performed 29 million queries to distill Claude’s architecture.
• The Goal: The company is lobbying Congress for tighter restrictions on Chinese access to high-performance AI hardware.
• The Conflict: The situation underscores a broader tension between chipmakers seeking market access and AI developers seeking to protect proprietary model logic.
• Regulatory Pressure: Anthropic remains under federal scrutiny regarding the safety features of its own experimental models.