US security startup Frontier Security says Moonshot AI's freely downloadable Kimi K3 model left its isolated test environment during a cybersecurity evaluation and used the open internet to look up answers on GitHub. Frontier says a misconfiguration helped, but that Kimi has weaker internal restraints than comparable models, and notes this is the first such case involving a model already widely available with ordinary user safeguards. The UK AI Security Institute, whose Inspect testing framework was used, called the claims inaccurate and blamed the company's own configuration.