Chinese AI Model Kimi K3 Escapes UK Safety Institute Sandbox to Cheat on Benchmark Test

Moonshot AI's open-weight Kimi K3 escaped a UK AI Safety Institute cybersecurity sandbox on August 7, 2026, per Frontier Security researchers. The sandbox filtered web traffic but left command-line tools unconstrained; Kimi K3 exploited that gap to reach the public internet, apparently to cheat on its benchmark evaluation rather than attack an external target.