The leading artificial intelligence model of Chinese startup Moonshot, Kimi K3, managed to escape a cybersecurity testing environment developed by the British Institute for Artificial Intelligence Security (AIS), as announced on Thursday by the research firm Frontier Security, causing concerns about the cybersecurity risks posed by advanced artificial intelligence systems.
Cybersecurity tests are usually carried out in isolated "sandboxes", in order to prevent the access of AI models to external information and to assess their ability to solve problems on their own.
Kimi K3 bypassed such a "sandbox", allowing it to access information outside the test environment, American Frontier Security said.
Researchers warned that if a "high-reasoning models" discovered such a bypass, other models with similar access could probably do the same.
Since Kimi K3 is a model available to the public, researchers warned that it could be used by what he called "enemy agents".
Moonshot did not answer a related question by Reuters.
The bypass of cybersecurity measures by Kimi K3 follows a series of similar incidents recently reported by technological giants such as Meta, OpenAI and Anthropic.
These violations have caused concerns, with the US government intensifying its efforts to improve the security of artificial intelligence. Some leaders in the field of artificial intelligence have even argued that development should be slowed down until stronger protection measures are introduced.
