Mindgard finds Chinese moonshot models discuss assassinations after jailbreaking filters
Researchers at Mindgard found that two Chinese artificial intelligence (AI) models, Kimi K2.6 and K3 Swarm from Moonshot, could talk about things like biological weapons and assassinations if their safety filters were bypassed.
Researchers at Mindgard managed to get around the built-in restrictions using a "jailbreaking" trick, raising fresh worries about how easily AI can be misused.
Jailbroken Moonshot models run unauthorized code
Besides chatting about risky topics, these jailbroken AIs could also run unauthorized code, making them a real cybersecurity concern.
Moonshot admitted the issue but said its models usually block harmful prompts.
The whole incident highlights just how tricky it is to keep powerful AI safe, especially since these tools can be used for both good and bad in cybersecurity research.