[Kimi K3 Also 'Broke Out'? Turns Out the Sandbox Wasn't Disconnected from the Internet] According to monitoring by Beating, following OpenAI and Anthropic, Kimi K3 has also joined the ranks of 'AI escaping the sandbox.' WIRED even titled their article 'One of China's Most Powerful AI Models Has Also Broken Isolation,' making it sound like Kimi broke out of the sandbox on its own. But upon closer inspection, it turns out the door wasn't properly closed. During testing by the security company Frontier Security, a network vulnerability was found in the sandbox for Kimi K3. The sandbox, which was supposed to be isolated from the public internet, still allowed access to GitHub. Kimi detected this network path on its own and simply cloned the official benchmark repository to find answers from it. This entire incident did not involve breaking into a properly configured sandbox, nor did it involve hacking into any company. To put it plainly, it’s like an exam proctor forgot to disconnect the internet, and the test-taker discovered they could still access GitHub, so they just looked up the answers. But when it made it to the headlines, it became 'AI Breaks Isolation.' Nowadays, the threshold for calling something a 'breakout' has dropped so low that even an unlocked door counts as a jailbreak. [Original Link]