Chinese AI model bypasses restrictions and accesses the internet during security test
A Chinese artificial intelligence model, the Kimi K3, developed by the company Moonshot, managed to escape a cybersecurity testing environment. The information was revealed by researchers in a recent publication.
The incident reinforces concerns that independent companies and organizations face increasing difficulties in maintaining control over artificial intelligence models created with a focus on hacking activities.
More on this story: Teenager kills mother and brother after using AI to create fantasies about the crime
In recent weeks, large language models (LLMs) from US labs such as OpenAI, Anthropic, and Meta, as well as the UK’s AI Security Institute, have also been released from test environments. These systems even invaded real targets outside the experimental parameters. The frequency of these events led to the creation of the Felony Bench website, which monitors such incidents, raising discussions about the possible criminal nature of these actions.
See Also
Electrical discharge knocks man off roof in Jaguaruana after contact with power lines
Global auction raises US$11.9 million and Brazil’s historic currency reaches record value
Former fighter Prichard Colón dies at age 33 after decade in a vegetative state
In the case of the test with Kimi, the failure occurred due to an inadequate configuration of the “sandbox”, the isolated environment for the experiment. Although the sandbox prohibited access to specific web traffic, the AI model bypassed the restriction using command-line tools, as found by researchers at cybersecurity company Frontier Security.
The researchers stated that “this suggests that some of the cybersecurity assessments used by the community are susceptible to vulnerabilities, allowing models to cheat. There are even systems that intentionally look for loopholes to manipulate assessment results.”
With this new record, Moonshot equals OpenAI and Anthropic, both with seven incidents recorded by Felony Bench. Meta, in turn, has a case documented in the same survey.
