February 23 (Reuters) - Anthropic, the developer of the artificial intelligence model Cloud, said in a blog post on Monday that three Chinese companies had exploited the program to illegally gain capabilities to improve their own models, and called for controls on the export of the electronic chips.

This follows a similar disclosure from OpenAI this month. A memo seen by Reuters indicated that OpenAI alerted US lawmakers that Chinese AI startup DeepSec was targeting OpenAI, the developer of ChatGPT software, and leading US AI companies to copy their models and use them in its own training.

Anthropic said that DeepSec, Moonshot and MiniMax interacted more than 16 million times with the Cloud program using approximately 24,000 fake accounts, in violation of the company's terms of service and regional access restrictions.

Anthropic explained that these companies used a technique known as "distillation," which involves an older, more established, and more robust AI model evaluating the quality of answers provided by a newer model, effectively transferring the knowledge of the older model.

The company added, "These campaigns are becoming increasingly intense and complex. The window of opportunity for action is limited, and the threat transcends any particular company or region."

Anthropic warned that models obtained through illicit "distillation" technology lack the necessary safeguards, posing serious national security risks. If these models are made available as open source, the risks are compounded as the capabilities proliferate freely outside the control of any government.

Anthropic, which raised $30 billion in its latest funding round and is currently valued at $380 billion, asserted that distillation attacks support the case for export controls, as restricting access to chips weakens the direct training capabilities of models and limits the scope of illicit distillation activity.

Anthropic reported that Operation DeepSec targeted inference capabilities across various tasks, creating secure, unmonitored alternatives, while Operation Moonshot focused on automated inference and the use of tools, as well as programming and data analysis.

MiniMax aimed to focus on automated programming, tool use, and formatting.

DeepSec, Moonshot and MiniMax have not yet responded to requests for comment.