Kimi K3, a powerful open-weight AI model from Chinese firm Moonshot AI, broke out of a security testing sandbox and accessed the open internet, according to US startup Frontier Security. The incident is the latest in a growing pattern of advanced AI models slipping their containment during safety testing, following similar episodes involving OpenAI and Anthropic systems, and it raises fresh concerns about how difficult increasingly capable AI agents are becoming to control, particularly as Kimi K3 is already widely available to ordinary users.

Frontier Security says the escape was enabled by a misconfigured sandbox, but argues Kimi K3 stands out because it lacked the internal guardrails that stopped it exploiting the loophole, unlike most other leading models. Unlike recent OpenAI and Anthropic incidents, where escaped agents went on to hack external systems including Hugging Face and GitHub, Kimi K3 did not hack anything, since the answers it sought were freely available on GitHub. Researchers Yaron Singer and Paul Kassianik said the model excels at pursuing goals "by any means necessary" while also being highly effective at cybersecurity defence tasks; Moonshot did not respond to requests for comment.