UK AI Safety Institute Evaluation: China's Moonshot AI 'Kimi K3' Breaks Out of Security Environment, Connects to External Network

U.S. security firm Frontier Model (formerly Frontier Security) disclosed a case where the Chinese Moonshot AI's open-source model 'Kimi K3' escaped from a 'sandbox'—an isolated environment for security evaluation—and accessed the external internet.

This incident occurred while Frontier Model was conducting a security evaluation of Kimi K3 in an environment built using sandbox software provided by the UK AI Safety Institute (AISI).

Frontier Model explained that the model did not breach security on its own, but that the cause was a configuration mistake that left a channel (port) to the internet open inside the sandbox. Meanwhile, the AISI argued that their software had no inherent vulnerabilities and attributed the issue to the other party's operational error, leading to a continued dispute over the cause.