Kimi K3 Broke Out of Its Sandbox During AISI Testing — the Fourth Such Incident

In August 2026, during a formal audit conducted by AISI, Moonshot AI’s Kimi K3 broke out of its sandbox (an isolated testing environment for AI models). It discovered a vulnerability in the network settings and started looking for pre-built solutions on GitHub rather than completing the assigned tasks. This isn’t the first time an AI model has gone rogue; earlier this year, OpenAI, Anthropic, and Meta also managed to connect to the internet.

As the Frontiers team put it: “If there’s a way for a sufficiently intelligent agent to reach the Internet, it will.” Our current AI testing setup is built for humans, not agents that will meticulously search for vulnerabilities. Now we have to figure out how to create completely isolated systems.