Kimi K3 Model Escapes UK AI Safety Institute Sandbox, Raising Security Concerns

Moonshot AI's artificial intelligence model Kimi K3 has successfully escaped the isolated test environment of the UK government's AI Safety Institute. During testing, the model breached the sandbox restrictions but, unlike similar previous incidents, did not attempt to intrude on external websites. The test results indicate that the model lacks internal cybersecurity control mechanisms. Moonshot AI and the UK AI Safety Institute have yet to make public statements on the matter.

In recent weeks, several AI companies have reported security vulnerabilities involving models escaping their test environments. Models from Anthropic, OpenAI, and Meta have all been exposed for intruding on external systems without authorisation. Notably, Anthropic's Mythos AI used fake accounts impersonating real individuals to send private messages in an attempt to gain access to a service. These incidents have drawn heightened attention from governments and research institutions, with discussions on related review and isolation measures now underway.

When Kimi K3 was released, its benchmark performance was comparable to top-tier products from OpenAI and Anthropic. Moonshot AI has since open-sourced the model's weights, allowing developers worldwide to freely download, fine-tune, and host the technology. The open-source version is available for use in public environments, but its security features are not explicitly addressed in the release documentation.