Security Researchers Say Chinese AI Models Could Be Coaxed Into Giving Bioweapons Guidance

A cybersecurity firm says it discovered that artificial intelligence models developed by Chinese AI company Kimi could be manipulated into providing guidance on creating bioweapons, according to the BBC.

The firm, Mindgard, said it identified the vulnerability in July, finding that two of Kimi's models, K2.6 and K3 Swarm, could be induced to evade the safety limits their developer had put in place to prevent the models from producing dangerous content.

How the safety limits were bypassed

AI developers typically build in restrictions intended to prevent their models from responding to requests for information that could cause serious harm, including instructions related to chemical, biological, radiological or nuclear weapons. Mindgard's research reportedly found ways to work around these restrictions in the Kimi models tested.

The findings raise broader concerns about the robustness of safety measures across AI systems, particularly as increasingly capable models are developed and released by companies around the world, including in China, where regulatory oversight of AI safety may differ from that in other countries.

The disclosure adds to a growing body of research examining whether AI safety guardrails, often implemented through content filtering and fine-tuning, can reliably prevent misuse when users employ adversarial techniques designed specifically to circumvent them.

It was not immediately clear from available reporting whether Kimi's developer has addressed the vulnerabilities identified by Mindgard or commented publicly on the findings.

The case underscores ongoing debate among researchers, policymakers and AI companies about how to balance rapid development of powerful AI capabilities with adequate safeguards against potential misuse for serious harm.