TL;DR

Four people who co-founded Z.AI, the Beijing-based maker of the GLM open-weight AI models, spent close to four years on a project part-funded by CGTRI, the China General Technology Research Institute, according to Bloomberg reporting relayed by Startup Fortune. MI5 publicly named CGTRI on 30 September as having “very strong ties” to the Ministry of State Security in Beijing. Bloomberg’s report, published on 19 October, makes no claim that GLM models themselves leak data.

What Bloomberg found

According to Startup Fortune’s account of the paywalled report, the work began in January 2018 and finished in December 2021, so it was still running when Z.AI was set up in 2019. Li Juanzi, who teaches computer science at Tsinghua University and co-founded Z.AI, led it, with money from CGTRI and China’s National Natural Science Foundation. Its stated goal was AI that scans social media posts to spot major events and map links between people.

In Startup Fortune’s reading, that matches what MI5 warned about: CGTRI money flowing into research whose results end up helping Chinese intelligence. When MI5 named CGTRI, the security service said the institute’s main purpose is funding academic research that improves the Ministry of State Security’s technical capability for spying, and that upwards of 100 academics based in Britain had worked on research it funded. China’s embassy in London rejected the allegations as “imaginary and purely fabricated”.

A question about people, not code

Startup Fortune draws the distinction that matters for businesses: the report concerns who built the models and who paid for their early research, not evidence that the models send data anywhere. It calls the link between that 2018-2021 work and the trustworthiness of models released years later “a genuinely open question”.

Looking forward

GLM models appeal because they are open-weight, can be run on a company’s own hardware and are cheap to use, and some development teams have built pipelines around them. In our view, the effect on UK organisations will come through procurement questions rather than any ban, and firms using GLM or other Chinese open-weight models should at least record the decision and the reasoning now, because security teams and public-sector clients are the parties most likely to start asking after an MI5 alert. Running a model on-premises limits data exposure, but it does not answer questions about provenance.