Anthropic has accused Chinese AI firms of stealing data from its Claude chatbot to enhance their models, raising security concerns.
Anthropic has leveled serious accusations against three Chinese AI startups—DeepSeek, Moonshot, and MiniMax—for allegedly engaging in extensive data theft to bolster their own AI models using Claude, Anthropic's chatbot.
According to the claims, these companies generated over 16 million interactions with Claude through around 24,000 fraudulent accounts, violating both usage agreements and geographic restrictions.
“We are confident that we can link each campaign to a specific firm based on IP address correlations, request metadata, and confirmations from industry partners. They specifically targeted Claude's unique features: agent reasoning, tool usage, and programming,” Anthropic stated.
The accused labs reportedly employed a technique known as distillation, which involves training a less powerful neural network on the outputs of a more advanced one. While distillation is a common and legitimate method used by AI labs to create efficient models, Anthropic warns that it can also be misused.
“Competitors can enhance their models at the expense of others' LLMs with significantly less time and cost compared to developing their own,” the company noted in a blog post.
Anthropic pointed out that the repercussions of such data theft extend beyond individual firms, stressing the need for swift action from the AI community, regulators, and industry stakeholders to combat this issue.
The Dangers of Distillation Abuse
Anthropic outlined the potential risks associated with illegally distilled models, which may lack essential safety measures. This poses national security risks, as American companies are focused on preventing the misuse of AI in areas such as biological weapon development and cyberattacks.
Models created through illegal means may integrate unregulated capabilities into military and intelligence operations, enabling authoritarian regimes to exploit advanced AI for harmful agendas, according to Anthropic.
Proposed Solutions
To address these concerns, experts at Anthropic advocate for export restrictions to protect U.S. leadership in AI. They argue that distillation attacks undermine these efforts, allowing foreign labs to gain technological ground.
“Without transparency regarding these attacks, the rapid advancements of Chinese labs could be wrongly viewed as evidence that export restrictions are ineffective,” the company added.
Anthropic has outlined several countermeasures, including:
- Enhancing detection for distillation patterns;
- Sharing technical indicators with other AI labs and cloud providers;
- Strengthening verification processes for educational and research accounts;
- Implementing measures to reduce the efficacy of illegal distillation.
This incident is not the first of its kind; previously, DeepSeek-R1 was suspected of stealing data from OpenAI shortly after its launch in January 2025.
Ongoing Tensions with the Pentagon
Dario Amodei, CEO of Anthropic, is set to meet with Defense Secretary Pete Hegseth at the Pentagon to discuss the military's use of Anthropic's AI models. There have been ongoing tensions, with Anthropic opposing the application of AI for mass surveillance and autonomous weapons development, while the Pentagon insists on using LLMs in all lawful scenarios.
This discord has led to discussions about potentially terminating Anthropic's contract with the Department of Defense.
Impact on Cybersecurity Sector
Following the launch of Claude Code Security, an AI vulnerability scanner by Anthropic, shares of major public cybersecurity companies experienced declines. The tool analyzes code for vulnerabilities, identifying critical issues that have persisted for years.
Recent reports indicate that the five largest publicly traded IT security firms in the U.S. have seen their stock prices drop significantly:
- Palo Alto Networks — down 14%;
- CrowdStrike — down 18%;
- Fortinet — down 12%;
- Cloudflare — down 18%;
- Zscaler — down 19%.
Analysts from Wedbush suggest that the market's reaction is overblown, asserting that these companies will demonstrate their value by 2026.
In a related development, OpenAI and Paradigm had previously introduced EVMbench, a benchmark for evaluating AI agents' abilities to identify and rectify vulnerabilities in smart contracts.
