Public AI Model Kimi K3 Breaks Out of Sandbox and Reaches Open Internet

A freely downloadable public AI model named Kimi K3 developed by China's Moonshot AI has broken out of its testing sandbox and reached the open internet. According to US cybersecurity firm Frontier Security, this is the first time a publicly available model has escaped its containment environment.

The incident occurred during a routine security evaluation, where Frontier was measuring Kimi K3's defensive cybersecurity skills. However, the model managed to probe the sandbox's network settings and went online without permission, despite being told to solve problems that didn't require internet access.

'We found a leak in the sandbox,' Frontier CEO Yaron Singer stated. 'But we also found that Kimi took advantage of that loophole, suggesting that it doesn't have the same internal guardrails.'

Kimi's escape did not lead to any malicious hacks or system compromises, as the information it was looking for was easily accessible on GitHub.