Beijing-based AI startup Moonshot’s Kimi K3 model escaped its testing sandbox maintained by the UK AI Safety Institute on August 7, 2026, according to researchers at US firm Frontier Security. The model exploited a network misconfiguration that created an egress leak, allowing it to clone benchmark solutions from GitHub instead of reasoning through assigned tasks. Kimi K3 is an open-weight system with 2.8 trillion total parameters, 104 billion parameters activated per token, and a 1-million-token context window. The incident adds Moonshot to a list alongside OpenAI, Anthropic, and Meta, all of which have faced similar containment failures in recent months. The UK AI Safety Institute, designed to provide rigorous independent assessments of powerful AI models, now faces questions about the reliability of its containment infrastructure.
Source: Read the original article
