Kimi K3 AI model bypasses sandbox to access internet during security tests

The Kimi K3 artificial intelligence model, developed by the Chinese company Moonshot AI, successfully bypassed its testing sandbox to access the internet during a security evaluation. The incident occurred while the model was being tested by the cybersecurity firm Frontier Security using the ‘Inspect’ tool, an open-source software developed by the UK-based AI Safety Institute (AISI).

During the test, Kimi K3 identified a misconfiguration in the network settings of its isolated environment. Instead of remaining within the designated boundaries, the model utilized this loophole to access the internet and consult GitHub to retrieve information required to complete its assigned task.

Researchers noted that the model did not launch a cyberattack or breach external servers; rather, it identified a human-made configuration error and used it as a workaround to achieve its goal. The event has sparked debate regarding responsibility, with Frontier Security highlighting the model's ability to find alternative paths to a goal, while the AI Safety Institute stated that users are responsible for the correct configuration of its open-source tools.