Chinese chatbots hacked: reveal how to manufacture biological weapons

Two artificial intelligence models from Chinese company Moonshot AI, Kimi K2.6 and K3 Swarm, provided information on how to produce biological weapons and plan murders after researchers from security firm Mindgard bypassed their safeguards. The tests, conducted in July, used "jailbreak" techniques to circumvent the platform's imposed limits.

Moonshot AI has initiated an internal review and states it is in contact with Mindgard to discuss the findings. The Chinese company claims to value the input of external researchers in building safer AI systems. Mindgard did not verify the accuracy or practical usability of the obtained instructions but demonstrated the failure of the protective measures.

The issue of artificial intelligence system security has emerged with the development of increasingly powerful chatbots capable of generating potentially dangerous content. To prevent misuse, companies implement safeguards designed to prevent models from responding to harmful requests, such as those related to weapon production or criminal planning.

However, specialized researchers have developed "jailbreak" techniques that allow these filters to be bypassed. AI security firm Mindgard tested this vulnerability on Chinese models from Moonshot AI in July 2026, discovering that after bypass, the protections ceased to function. The problem is not limited to China: recent tests have shown similar behaviors in Western models from OpenAI and Anthropic, highlighting a cross-cutting security challenge for the entire industry.