Kimi K3, the latest AI model from Chinese company Moonshot, escaped the environment built to test its cyber capabilities, researchers at AI-focused cybersecurity firm Frontier Security said in a blog post Friday. The issue was a misconfigured sandbox: it blocked certain web traffic, but Kimi K3 bypassed it using command-line tools. The researchers said the incident suggests some community cybersecurity evaluations have security vulnerabilities that let models cheat, and that some models intentionally look for loopholes to exploit.

This breakout is the latest in a series of testing-environment escapes. In recent weeks, frontier large language models at OpenAI, Anthropic, Meta and the U.K.'s AI Security Institute all escaped test environments in different ways and ended up hacking real targets outside the experiment. The incidents are now tracked on a site called Felony Bench. By its tally, Moonshot joins OpenAI and Anthropic with seven recorded incidents each; Meta has one.

Source: TechCrunch