China's AI model Kimi K3 exits UK government sandbox

JAKARTA - The Chinese artificial intelligence model, Kimi K3, is reported to be able to get out of the British government's cyber testing environment. Researchers then highlighted the safety system of the AI model made by Moonshot AI.

Anadolu Agency, quoted on Friday, August 7, reported that Kimi K3 was tested in a sandbox operated by the UK's AI Security Institute.

A sandbox is a controlled environment designed to safely test AI systems.

The US cybersecurity research company, Frontier Security, said Kimi K3 managed to find a way out of the environment.

However, the model did not try to break into external websites.

The researchers said the incident showed a weakness in Kimi K3's cyber security system.

"Kimi's model, which is available to the public, does not have such a safety system," said Frontier Security founder and CEO Yaron Singer to Bloomberg, quoted by Anadolu Agency.

Similar cases have previously been reported involving AI models from Anthropic, OpenAI, and Meta. According to researchers, some of these models have accessed the organization's systems outside the testing environment.

Kimi K3 previously attracted attention because its test results were able to rival a number of leading models from OpenAI and Anthropic.

Moonshot AI also released the weights of the Kimi K3 model publicly. That way, developers can download, modify, and run the model independently.