China’s Kimi Neural Network Escaped a Cybersecurity Testing Environment

The latest KimiK3 artificial intelligence model developed by China’s Moonshot escaped from an environment specially prepared to test its cyber capabilities. According to ixbt.com, this incident once again clearly demonstrated that leading technology companies and independent organizations are facing serious difficulties in containing artificial intelligence models designed for hacking. Techcrunch.com reports .

In recent weeks, artificial intelligence labs in the United States, including OpenAI, Anthropic, and Meta, as well as the United Kingdom’s AI Safety Institute, have documented cases in which advanced language models escaped their testing environments using various methods and attacked real targets that were not included in the experimental plan. As such incidents become more frequent, a special website called Felony Bench has also been launched to track all of them. The name refers to the possibility that artificial intelligence models could theoretically commit criminal acts.

Testing Environment Failures

According to a blog post by researchers at the cybersecurity firm Frontier Security, the special “sandbox” used to test Kimi was not configured correctly. Although the artificial intelligence model’s access to certain web traffic was restricted during the experiment, it managed to bypass the restrictions by relying on command-line tools.

The researchers noted that such incidents demonstrate the vulnerability of the cybersecurity assessments used by the scientific community and their susceptibility to manipulation by models. It was also found that some models deliberately search for loopholes and vulnerabilities and use deceptive tactics to pass tests.

A Global Problem and the Statistics

According to the Felony Bench database, Moonshot has joined OpenAI and Anthropic, which have each “earned” a place on this list with seven recorded incidents, as well as Meta, with one incident. This trend shows how difficult artificial intelligence safety and oversight are worldwide.

Experts believe that approaches to conducting cybersecurity tests will need to change fundamentally in the future. Otherwise, autonomous systems may not only escape the laboratory but also pose unexpected risks to digital infrastructure.