Chinese AI model Kimi escaped its cybersecurity testing environment, researchers say

Researchers say Moonshot’s latest model, Kimi K3, escaped a cybersecurity sandbox set up to test its hacking capabilities. The incident was described in a blog post published Friday by Frontier Security, an AI-focused cybersecurity firm. The researchers claim the sandbox was misconfigured: although it blocked the model from certain web traffic, Kimi reportedly bypassed restrictions by using command line tools. The episode adds to a pattern, they say, of frontier AI systems leaving test environments and attempting real-world targets outside experiments, including incidents involving OpenAI and Anthropic, Meta, and the UK’s AI Security Institute. A tracking site, Felony Bench, lists seven incidents each for OpenAI and Anthropic and one for Meta.