Logo of the Chinese startup Moonshot AI next to the Kimi application displayed on a cell phone screen, in an illustrative image from July 24, 2026. REUTERS/Florence Lo/Illustration/File Photo/File Photo The Kimi K3, the main artificial intelligence model of the Chinese startup Moonshot, managed to escape from an isolated environment used for digital security tests, according to the research company Frontier Security reported this Thursday (6). ??Do you have any reporting suggestions? Send to g1 During this type of testing, AI systems often run in "sandboxes" — isolated environments that prevent access to external information. The objective is to evaluate the extent to which these models can solve problems on their own, without interference from the real world. According to Frontier Security, an American company specializing in cybersecurity, the Kimi K3 managed to bypass one of these "sandboxes", accessing information that should have been blocked during the test. The researchers claim that if a "high-reasoning model" was able to find this shortcut, other systems with similar characteristics and equivalent access could also do the same.
Chinese AI model finds breach in security test and escapes isolated environment
Logo of the Chinese startup Moonshot AI next to the Kimi application displayed on a cell phone screen, in an illustrative image from July 24, 2026. REUTERS/Florence Lo/Illustration/File Photo/File Photo The Kimi K3, the...
As the Kimi K3 is publicly available, researchers warn that it could be exploited by "malicious actors", which would make the incident potentially more serious. Moonshot did not immediately respond to Reuters' request for comment. The episode adds to a series of similar incidents recently reported by companies including Meta, OpenAI and Anthropic, raising concerns about the security of increasingly sophisticated artificial intelligence models.
This story was originally published by G1. Visit the original publication for further details.
Open original publication
