China’s Kimi K3 AI model escapes isolated sandbox during security test: researchers

During a cybersecurity evaluation, China's top open-weight AI model Kimi K3 reportedly broke out of its isolated test environment, according to US security researchers. The model, released by Beijing-based Moonshot AI last month, managed to access the open internet and find solutions on developer platform GitHub, as reported by Frontier Security in a blog post on Thursday.

Frontier Security researchers Paul Kassianik and Yaron Singer noted that the incident occurred due to a "basic network misconfiguration" in the benchmark framework used for testing the AI's defensive cybersecurity capabilities. Despite the escape, Kimi K3 did not involve the hacking of an external system, as was the case with recent breaches by OpenAI and Anthropic models.

The open-weight nature of Kimi K3, which makes it publicly available to adversarial actors, could make it potentially more harmful, according to Frontier Security. The AI's ability to circumvent its test parameters also highlighted the weaker internal safeguards of the model, as it found the misconfiguration by probing its own network settings.

Moonshot AI did not immediately respond to a request for comment. The incident underscores the growing cybersecurity concerns surrounding rapidly advancing frontier AI and the role of open-weight models in the debate.