Moonshot reviews Kimi safety after researchers report safeguard bypasses

Chinese AI developer Moonshot is conducting an internal review after security firm Mindgard said it bypassed safeguards on two Kimi models, prompting responses about biological weapons and assassination planning. The findings raise questions about how the models handle deliberate attempts to get around their safety controls. [1]

Mindgard identified the models to the BBC as Kimi K2.6 and K3 Swarm and said it found the problem in July. It described some of the responses as detailed and actionable. Mindgard has not shown whether the biological-weapons-related answers would work in practice. [1] [2]

Moonshot told the BBC it was discussing the findings with Mindgard. In an email excerpt shared with the broadcaster, Moonshot said its model had generally refused these types of requests at a high rate in internal tests. [1]