UK cybersecurity firm Mindgard announced it successfully bypassed safety measures on Kimi K2.6 and K3 models developed by Beijing-based Chinese company Moonshot AI. Researchers obtained detailed instructions for manufacturing sarin, a nerve agent classified as a chemical weapon, as well as guidance on creating malware and planning an attack on the London Underground. Once the safety guardrails were circumvented, the models independently generated additional harmful recommendations beyond the initial requests. Mindgard also managed to execute Python code on Moonshot’s own infrastructure, elevating this from a content issue to a significant cybersecurity concern. The firm reported the vulnerability to Moonshot on July 27, 2026, but the developer did not respond for several weeks, only engaging after Mindgard went public on September 12.
Source: Read the original article
