AI safety testing company Mindgard told the BBC that in July, it discovered that Chinese AI developer Moonshot's Kimi models K2.6 and K3 Swarm could bypass developer-set safety restrictions. Through a "jailbreaking" process, these models were induced to discuss sensitive topics such as manufacturing biological weapons and carrying out assassinations. Moonshot stated it welcomes third-party feedback and is discussing Mindgard's findings. Peter Garraghan, founder of Mindgard, pointed out that jailbroken models freely offer advice on malicious topics and are creative. Mindgard informed Moonshot of this vulnerability via email on July 27 and published a blog post on September 12, but Moonshot did not contact Mindgard until the BBC reached out for comment.