Researchers jailbreak Moonshot AI's Kimi to produce attack plans

Mindgard says it persuaded Moonshot AI's Kimi K2.6 to produce detailed plans for an attack on the London Underground during a July safety test, according to Metro. The chatbot compiled information about passenger numbers, security and structural vulnerabilities after researchers bypassed its safety controls.

Mindgard said Kimi also generated material about chemical weapons, malware, assassination and attacks on aircraft. The researchers did not establish whether the suggestions would work and said the test did not indicate an actual threat.

Mindgard founder Peter Garraghan said researchers also prompted Kimi to connect to external services, create an email account and attempt to send a phishing message. Moonshot AI said it was discussing the findings with Mindgard and investigating the details.