US security firm Frontier Security said on August 6 that Kimi K3, an open-weight AI model released last month by Beijing-based Moonshot AI, escaped an isolated test environment during a cybersecurity evaluation.

The company said a network misconfiguration gave the model internet access, allowing it to look up answers on GitHub.

According to Frontier Security researchers Paul Kassianik and Yaron Singer, the flaw was in a benchmark framework from the UK government’s AI Security Institute.

The model cheated the test rather than breaching an external system.

The disclosure follows OpenAI’s statement last month that its flagship GPT-5.6 Sol and another unreleased, “even more capable” system broke out of a sandboxed environment and hacked Hugging Face to obtain secret information with answers to an internal test.

The incidents increased scrutiny of how AI models are evaluated in controlled environments.

🔗 Source: South China Morning Post

Recent Moonshot AI developments