"Hacker AI": A Neural Network from China That Impressed the Market Sped Out of Control During Testing

Bloomberg noted that the latest artificial intelligence model, Kimi K3, developed by the Chinese startup Moonshot, was able to escape the isolated test environment of the British government’s AI Security Institute.

The incident was reported by Frontier Security, an American cybersecurity research firm. According to its founder and CEO, Yaron Singer, the publicly available Kimi model lacks the necessary security restrictions. “Essentially, this makes it a very good model for hackers,” he told Bloomberg.

Although, unlike the models from Anthropic, OpenAI, and Meta—which were also able to independently access the internet during recent tests—Kimi K3, once connected to a large network, did not attempt to attack third-party websites or systems. Nevertheless, the researchers point out that the lack of controls poses a threat in and of itself.

A Moonshot spokesperson did not immediately comment on the situation. The AI Security Institute, from whose test environment Kimi K3 managed to escape, also did not respond to inquiries.

Thus, Moonshot has become the fourth AI lab whose models have gone beyond test environments. Previously, similar incidents were recorded during tests of systems developed by the U.S. companies Anthropic, OpenAI, and Meta. However, once online, their models also hacked into third-party organizations’ systems. These incidents have raised concerns among researchers and government officials regarding security and the ability of developers to control their AI models.

This article is being updated

This article was AI-translated and verified by a human editor