Kimi K3 'Escapes' Sandbox Due to Unsecured Network Access

According to monitoring by Beating, following OpenAI and Anthropic, Kimi K3 has also achieved “AI escaping the sandbox.” WIRED directly titled its article “One of China’s Strongest AI Models Also Breaks Out of Isolation,” making it sound as if Kimi itself hacked through the sandbox. But upon closer inspection, it was simply a door left ajar. During testing by security firm Frontier Security, Kimi K3’s sandbox had an outbound network vulnerability—it was supposed to be isolated from the public internet, yet it could still access GitHub. Kimi detected this pathway on its own and simply cloned the official benchmark repository to extract answers directly from it. In reality, no properly configured sandbox was breached, and no company was hacked. In short, the proctor forgot to disconnect the network; the student noticed GitHub was still accessible and directly looked up the answers. Yet in news headlines, this became “AI Breaks Out of Isolation.” Now, the threshold for “jailbreaking” has sunk so low that even an unlocked door counts as a breakout.