AI Agent Provides Researchers With Instructions to Create Biological Weapons

The researchers managed to convince two AI models of Kimi Chinese company Moonshot to tell how to create biological weapons and commit murder. The company then began an internal audit. This was told to the BBC by experts of Mindgard, which conducted tests of the AI agent.

According to Mindgard experts, in July, the Kimi K2.6 and K3 Swarm AI models bypassed the security restrictions set by the developers. This was discovered during a process known as “jailbreaking.” The researchers tested whether the AI ignored built-in defense mechanisms from providing potentially dangerous information. In the end, they were able to obtain such information.

Moonshot representatives have already said that they welcome the participation of third-party specialists in the work on AI agents. They consider this to be “the most important condition for creating better and safer artificial intelligence systems.”.

Eugene Hvostik