According to DynaBeat Monitoring, following OpenAI and Anthropic, Kimi K3 has also successfully achieved the "AI Sandbox Escape." WIRED directly titled the article "One of China's Most Powerful AI Models Also Breaks Isolation," making it sound like Kimi himself personally broke out of the sandbox.
However, upon closer inspection, it was another case of the door not being properly closed. During Frontier Security's testing of Kimi K3, a sandbox had an outbound network vulnerability. While it should have been isolated from the public network, it was still able to access GitHub. Kimi discovered this network path and promptly cloned the official benchmark repository to find answers.
This incident did not involve breaching a properly configured sandbox or hacking into any company's systems.
In simpler terms, it was akin to an exam proctor forgetting to disable the internet, allowing the test taker to access GitHub directly for answers. This was then sensationalized in the news headline as an "AI Breakout." The threshold for "jailbreaking" has now reached a point where even a door left unlocked is considered an escape.
