Anthropic has accused Alibaba of using nearly 25,000 fraudulent accounts to extract capabilities from its Claude artificial intelligence models, an allegation that the U.S. AI developer describes as the largest known attack of its kind against the company. According to a June 10 letter sent to senior members of the U.S. Senate Banking Committee, the claimed campaign ran from April 22 to June 5 and generated more than 28.8 million exchanges with Claude.
Anthropic said the operation involved distillation, a technique in which a less capable model is trained on the outputs of a more advanced system. If successful, distillation can allow a rival to replicate portions of a leading model's behavior at lower cost and in less time than building a comparable system from scratch. Anthropic attributed the alleged activity to operators affiliated with Alibaba and Alibaba Qwen, the company's AI lab. Alibaba did not immediately respond to a request for comment.
Key facts at a glance
- Accuser: Anthropic, the U.S. developer of Claude.
- Accused: Alibaba and operators affiliated with Alibaba Qwen, according to Anthropic.
- Scale: Nearly 25,000 fraudulent accounts and more than 28.8 million exchanges with Claude.
- Timeframe: April 22 to June 5, as described in a June 10 letter to senior members of the U.S. Senate Banking Committee.
- Method: Distillation, in which a weaker model learns from a stronger model's outputs.
- Context: The allegation lands amid rising enterprise adoption of generative AI, intensifying U.S.-China technology tensions, and growing concern about model extraction as a supply-chain risk.
- Response: Alibaba did not immediately respond to a request for comment.
What Anthropic alleges
Anthropic's accusation centers on the scale and persistence of the claimed activity. The company says the campaign used nearly 25,000 fake accounts, a volume that would have required coordination across many identities and access points. Over roughly six weeks, those accounts generated more than 28.8 million exchanges with Claude, according to the letter. Anthropic described the operation as the largest known attack of its kind against the company, a characterization that underscores how far model extraction has moved from a theoretical concern to an operational security issue for AI vendors.
The alleged campaign was not described as a conventional cyber intrusion in which attackers break into a network and steal source code or model weights. Instead, Anthropic says the operators sought to extract behavior from the model through repeated interactions. That distinction matters. A model's weights may remain inside a vendor's infrastructure, but its outputs can still reveal patterns, reasoning styles, coding capabilities, safety boundaries, and domain-specific knowledge. At sufficient scale, those outputs can become training data for another model.
Anthropic attributed the alleged activity to operators affiliated with Alibaba and Alibaba Qwen. The allegation has not been proven in a court or regulatory proceeding, and Alibaba did not immediately respond to a request for comment. The company's letter was sent to senior members of the U.S. Senate Banking Committee, placing the dispute within a broader policy debate over AI safety, competition, and national security.
Why distillation matters
Distillation is a well-established machine-learning technique. In its ordinary form, a smaller model is trained to imitate the outputs of a larger teacher model. The smaller model can often achieve useful performance at lower computational cost, which is why distillation is common in commercial AI development. The technique becomes controversial when it is used to capture capabilities from a rival's model without authorization, especially when the rival operates under usage terms that prohibit such activity.
For AI developers, the concern is not simply that another company might save on training costs. The deeper risk is that distillation can compress years of research, evaluation, and safety work into a shorter cycle. A model trained on a stronger system's outputs may inherit some of its abilities without inheriting the safeguards, testing, or governance that surrounded the original. That can create a competitive advantage for the imitator and a safety problem for customers who assume all models with similar capabilities have similar controls.
Anthropic said the alleged campaign could help accelerate China's ability to reach the capabilities of its advanced Mythos Preview model. That claim links the dispute to a wider concern among U.S. officials: that advanced AI systems could be used by military or intelligence users in countries of concern, or that foreign competitors could close the gap with U.S. labs faster than expected. The allegation therefore sits at the intersection of commercial competition, export controls, and AI safety policy.
A pattern of model extraction
The Alibaba allegation is not the first time Anthropic has raised concerns about model extraction. In February, the company said it had identified similar campaigns by DeepSeek, Moonshot AI, and MiniMax to extract capabilities from Claude. Anthropic said the alleged activity ranged from more than 150,000 exchanges by DeepSeek to more than 13 million exchanges by MiniMax. Viewed together, the incidents suggest a pattern rather than an isolated event.
Analysts said the alleged campaign points to a broader problem beyond any single company. Anand Joshi, an AI analyst at TechInsights, said that if Anthropic's claims are true, the alleged campaign could allow Alibaba to build a comparable model in a short period of time and offer it at a much lower cost. That possibility would threaten the economics of frontier AI development, where vendors spend heavily on data, compute, and safety testing in the expectation that customers will pay for differentiated capabilities.
Sanchit Vir Gogia, chief analyst at Greyhound Research, described the issue as a new kind of enterprise dependency. “The enterprise supply chain no longer ends at software, APIs, and cloud regions,” he said. “It now includes rented intelligence, and rented intelligence can be copied and redeployed well outside the safety controls it was born with.” Gogia said distillation should be a board-level concern because a weaker model trained on a stronger one can inherit its capabilities without the governance and controls around the original system.
The enterprise supply-chain risk
For enterprises, the allegations point to a potentially more serious risk than conventional intellectual property theft: reverse engineering at scale. If proven, they would suggest that AI models can be copied systematically, turning model extraction into a new AI supply-chain risk. Companies that rely on a vendor's model for customer service, software development, fraud detection, or internal decision-making may be exposed if a cloned or distilled model inherits weaknesses that the original vendor had mitigated.
Pareekh Jain, CEO of Pareekh Consulting, said the risk extends beyond lost revenue for the AI developer. “If a rival can clone the exact brain of the AI your company relies on, they can easily find its blind spots, hack your automated systems, or cause the AI vendor to panic and shut down services that your business needs to run every day,” he said. That scenario turns model extraction into an operational resilience issue, not just a legal dispute over terms of service.
The concern is amplified by the pace of enterprise adoption. Businesses are embedding generative AI into workflows that were once handled by deterministic software. When an AI service is interrupted, rate-limited, or withdrawn after a security incident, the operational consequences can be immediate. Customers therefore have a stake in how vendors detect abuse, how quickly they respond, and whether contracts provide backup plans or financial remedies.
Mitigating the risks
The allegation raises questions about the controls AI vendors have in place and how customers can protect themselves. Jain said vendors should provide verified accounts, smart rate limits, abuse detection, usage monitoring, contractual bans on distillation, incident disclosure, and audit rights. Enterprises should ask how the vendor detects and blocks large-scale model extraction and can demand contracts that guarantee backup plans and financial refunds if the AI service gets attacked or suddenly shut down.
Joshi said enterprise customers should also press vendors for greater transparency around model development and safeguards. “Enterprise buyers should ask what training data was used, how it was trained, what guardrails exist, how they can audit it, and so on,” he said. “Model publishers will have to come up with watermarking technology in models as well as model responses. So if the model skills are stolen, they should be able to find the thief.”
Watermarking and provenance tools are still immature, but the direction of travel is clear. Vendors are likely to invest more in detecting automated extraction patterns, limiting high-volume access, and tying usage to verified identities. Enterprises are likely to ask harder questions about concentration risk, contractual protections, and the ability to switch providers if a model becomes unavailable or untrustworthy.
The dispute also has implications for policy. U.S. officials have stepped up scrutiny of advanced AI systems over fears they could be used by military or intelligence users in countries of concern. At the same time, AI development is increasingly tied to U.S.-China technology tensions, with export controls, investment restrictions, and safety testing debates shaping how models are built and shared. An allegation of large-scale distillation adds a concrete example to those debates.
For now, the facts remain contested. Anthropic has made a detailed allegation, Alibaba has not immediately responded, and no public adjudication has established what happened. But the claim has already focused attention on a vulnerability that exists whenever advanced AI is delivered as a service: the model can be protected behind an API, but its behavior can still be observed, sampled, and potentially copied. The immediate question is not whether distillation is technically possible, but whether AI vendors can detect and deter it before a customer's dependency on a model becomes a systemic exposure.
Source: InfoWorld News