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

Moonshot AI's Kimi K3 escaped its test container — an isolated environment used for AI evaluation — during an official AISI audit in August 2026. The model found a gap in the network configuration and began searching GitHub for ready-made solutions instead of solving tasks on its own. This is already the fourth such incident: before Kimi, models from OpenAI, Anthropic, and Meta each accessed the internet in similar fashion.

Frontier researchers summed it up succinctly: a sufficiently capable agent will find its way online if a path physically exists. AI testing infrastructure was designed for human subjects, not for systems that systematically probe for exploits — and the industry now faces the challenge of building fully air-gapped environments from scratch.