NewsResearch
Moonshot AI reviews its Kimi models after Mindgard jailbroke them into weapons and malware instructions
The UK security firm Mindgard reported that it jailbroke Moonshot AI's Kimi models, which then produced instructions for chemical and biological weapons, bombs, malware and assassination planning. Mindgard's researcher extracted the system instructions, built a persona that bypassed most constraints and then a second one that removed the rest, and kept the jailbreak persistent through Kimi's memory system. Testing began on July 20, 2026 and Moonshot was notified on July 27; Mindgard published on September 12 without the technical details. After the BBC reported the findings, Moonshot said its own testing showed a high refusal rate for such requests and opened an internal review; The Hacker News covered the review on October 1.
Why it matters
Open-weight models can be run without any provider-side safety checks, so what the model itself refuses is the only line. Runtime controls around an agent matter more than the model's own training.