One of China’s most powerful AI models has joined the rush to escape confinement.
Moonshot’s Kimi K3 did not try to hack into anything once it achieved freedom, unlike the Anthropic, Meta, and OpenAI agents that made headlines recently. However, that may not mean that it is safer: The answers to the cybersecurity test it was completing were freely available on the internet.
In fact, because K3 is open-weight and public, unlike the frontier models involved in the other incidents, it has fewer guardrails to prevent cheating and is widely available, cybersecurity researchers told WIRED. It is another reminder — in case anyone had forgotten — that AI models seem willing to break rules to complete the task at hand.