Bloomberg: Chinese AI model escapes UK cyber-testing environment

Chinese artificial intelligence model Kimi K3, developed by Moonshot, escaped a UK government cybersecurity testing environment, raising fresh concerns over AI safety and the ability of companies to control increasingly powerful models, Bloomberg reports, citing US cybersecurity research firm Frontier Security.

It said Kimi K3 managed to break out of a sandbox operated by the UK's AI Security Institute (AISI).

Although the model did not attempt to breach external websites, unlike several recent incidents involving US-developed AI systems, researchers made it clear that the test exposed a lack of built-in cybersecurity safeguards.

"Kimi's model, which is publicly available, does not have these guardrails in place," Frontier Security founder and CEO Yaron Singer told Bloomberg. "Basically that makes this a very good hacking model."

Moonshot did not immediately comment on the findings, while the AI Security Institute also did not respond to requests for comment.

The incident follows similar breaches reported by US AI developers Anthropic, OpenAI and Meta, whose models recently escaped testing environments and, in some cases, infiltrated the systems of external organisations, including AI platform Hugging Face. Those incidents prompted renewed calls from researchers and government officials for stricter AI safety testing and more secure evaluation environments.

Bloomberg noted that Kimi K3 recently drew global attention after achieving benchmark results comparable to leading models from OpenAI and Anthropic, marking a significant breakthrough for Moonshot, which has largely been overshadowed by Chinese rival DeepSeek.

The company has also released the model's weights, allowing developers to download, modify and host the system independently.

By Bakhtiyar Abbasov