While Anthropic has not publicly named the other two firms involved, sources familiar with the investigation suggest they are well-funded startups with close ties to China’s national AI initiatives. The company has filed formal complaints with U.S. authorities and is working with international cybersecurity partners to trace data exfiltration pathways. The incident comes amid heightened U.S. scrutiny over AI chip exports to China, as policymakers debate whether restrictions on advanced semiconductors are sufficient to curb technological advancement abroad.

According to MSNBC, the timing of the disclosure coincides with ongoing congressional hearings on the Export Administration Regulations (EAR), where lawmakers are considering expanding controls to include AI training data and model weights. "If you can’t export the chip, you export the data," one policy analyst told MSNBC. "This case shows that the battlefield has shifted from hardware to intellectual property."

Anthropic’s internal audit revealed that the fraudulent interactions were structured to maximize information yield: users submitted prompts designed to extract ethical boundaries, reasoning chains, and even latent safety mechanisms embedded in Claude’s architecture. Some queries mimicked academic research prompts, while others attempted to trigger model hallucinations or contradictions—techniques known in AI research as "model probing." The data harvested was reportedly used to fine-tune proprietary models that now compete directly with Claude in both Chinese and international markets.

"This undermines the entire premise of open, ethical AI development," said Dr. Elena Rodriguez, an AI ethics researcher at Stanford’s Institute for Human-Centered Artificial Intelligence. "If companies can freely mine models without consent, it creates a race to the bottom where transparency becomes a liability."

DeepSeek and the other implicated firms have not issued public statements in response to the allegations. However, industry insiders note that Chinese AI companies have increasingly relied on "data distillation" techniques—training smaller models on outputs from larger, proprietary systems—as a workaround for limited access to high-end computing resources.

Anthropic has since implemented new countermeasures, including rate-limiting based on behavioral analysis, CAPTCHA challenges for high-volume users, and cryptographic watermarking of model outputs to trace unauthorized reuse. The company is also exploring legal avenues under trade secret and copyright law, potentially setting a precedent for how AI-generated content is protected globally.

As the U.S. and China deepen their technological rivalry, this case underscores a new frontier in AI competition: not just who builds the fastest chips, but who can safeguard the intellectual DNA of their models. For now, Anthropic’s findings serve as a wake-up call to the entire AI industry: in the age of generative models, data is the new oil—and theft is becoming algorithmic.