Китайский разработчик искусственного интеллекта Moonshot проводит internal check after researchers from Mindgard, a company that tests the security of AI systems, managed to get two popular Kimi models to provide information on how to create biological weapons and how to carry out killings.

In Mindgard, told the BBC that as early as July its specialists identified a way to bypass safety restrictions in the Kimi K2.6 and K3 Swarm models.

The vulnerability was discovered during so-called “jailbreaking” — testing in which researchers, using a series of complex instructions, try to make the AI ignore restrictions imposed by the developers. According to Mindgard representatives, the protective mechanisms were supposed to prevent Kimi from discussing potentially dangerous topics.

Read more on our website.

🔗 Website without VPN YouTube Mailing New application BBC World Service