AI Distillation

Anthropic has raised fresh concerns about the security of advanced AI models after alleging a large-scale attempt to access Claude through nearly 25,000 fake accounts. The company said the activity generated more than 28.8 million conversations and was tied to actors affiliated with Alibaba and Alibaba Qwen. The allegations add context to the recent decision by the U.S. government to restrict access to Anthropic’s advanced models Mythos and Fable. For Anthropic, the issue is not only about unauthorized use, but also about the risk that powerful AI capabilities can be copied through distillation faster than expected. What Anthropic told the Senate In a letter dated 10 June and addressed to Senator Tim Scott and Senator Elizabeth Warren, Anthropic said it had detected what it described as an illicit operation aimed at Claude. The company sent the document ahead of a scheduled Senate Banking Committee hearing on artificial intelligence. According to Anthropic, the operation ran from 22 April to 5 June 2026. The company said the scale of the activity suggested a coordinated effort rather than isolated misuse. Alibaba had not responded publicly to requests for comment when the allegations became known. That left Anthropic’s claims without an immediate direct rebuttal from the company named in the letter. Why distillation matters The dispute centers on AI distillation, a technique that uses the answers of a more advanced model to train a weaker one. The approach can speed up development because engineers do not have to build every capability from the ground up. Anthropic argues that campaigns like the one it described could help China narrow the gap to models such as Mythos Preview. In that view, the threat is not just unauthorized access, but the possibility that leading AI systems can be replicated without the same level of research and investment. That concern helps explain why access to models such as Mythos and Fable has become politically sensitive. As AI systems become more capable, they are increasingly treated as strategic assets rather than ordinary software products. A pattern Anthropic says it has seen before This is not the first time Anthropic has pointed to suspected distillation activity involving Chinese AI firms. In February, the company said it had identified another campaign linked to startup DeepSeek and two other AI labs in China. Anthropic said the DeepSeek-linked operation involved more than 150,000 conversations with Claude. It also said Moonshot AI accounted for more than 3.4 million conversations, while MiniMax generated more than 13 million conversations. At the time, the company warned that such campaigns were increasing in “intensity and sophistication.” It said the problem required quick and coordinated action from both the AI industry and government. Washington is paying closer attention Anthropic said in its letter that it supports U.S. efforts to counter such activity. The company emphasized the importance of sharing threat intelligence and strengthening cooperation between private AI firms and the government. The latest allegations arrive as Washington takes a more cautious approach to Chinese access to leading AI systems. That caution is part of a broader debate over how the U.S. should guard advanced technology in a period of intensifying competition. Alibaba was added this month to a Pentagon list of Chinese companies said to be linked to the military, a designation the company is challenging. DeepSeek, by contrast, has not been placed on the U.S. trade blacklist, even though an interagency government committee reportedly views it as a national security risk. The differing treatment shows that U.S. policy on AI-linked Chinese companies is still evolving. What is clear, however, is that models like Claude, Mythos, and Fable are now central to a larger struggle over access, imitation, and strategic control. Source: www.indiatoday.in

Anthropic Reveals Suspected Fake-Account Campaign Targeting Claude, Alibaba Name Surfaces