Urgent investigation into Moonshot after Kimi models provided guidance on biological weapons and assassinations

Chinese artificial intelligence company Moonshot has launched an immediate internal investigation after researchers at security firm Mindgard obtained responses from two of its Kimi models on producing biological weapons and organizing assassinations.

According to statements Mindgard made to the BBC, the Kimi K2.6 and K3 Swarm models were able to bypass the safeguards put in place by their developers.

The experiment used a technique known as “jailbreaking,” which employs complex prompts to test whether an artificial intelligence system can circumvent safety restrictions. Those barriers were supposed to prevent Kimi from addressing matters of this kind.

Mindgard notified Moonshot of the problem on July 27 and published its test results in September. The company stressed that, in its assessments, the models generally rejected such requests to a large extent.

Moonshot told the BBC that it welcomes the involvement of third parties, calling it “a key pillar for building better and safer artificial intelligence.”

The case has become public as the artificial intelligence sector debates the safest path for future development: closed, proprietary systems such as ChatGPT and Anthropic’s Claude platforms, or open-source technologies.

Kimi is classified as an “open-weight” model. In principle, this means a user can obtain the model and run it on their own computing infrastructure.