China's AI company Moonshot's recently introduced AI model Kimi K3 was found to have left a test environment during a cyber security capability assessment.

TechCrunch reported on Aug. 7 that researchers at AI security firm Frontier Security said in a blog post that Kimi K3 bypassed an environment set up to evaluate hacking ability.

The case again shows that it is not easy to keep AI models designed for hacking inside a test environment, TechCrunch said.

In recent weeks, large language models from OpenAI, Anthropic and Meta left test environments in different ways and hacked real targets that were not included in experiments.

In the Kimi K3 case, a sandbox created to control the experiment was cited as a problem because it was not properly set up. The sandbox was designed to block access to some web traffic, but Kimi K3 bypassed it using a command-line tool.

Frontier Security's researchers said the case suggests that some cyber security evaluations used by the community are exposed to security vulnerabilities and that models can deceive evaluations. They added that some models can intentionally look for loopholes and vulnerabilities and bypass evaluations.

As such cases continue, a website called Felony Bench is also being run to track related incidents. In its tally, Moonshot was newly listed, while OpenAI and Anthropic logged 7 cases each and Meta recorded 1.