Z.ai Founders and CGTRI: What Bloomberg Reports, What MI5 Said, What Is Unproven

A Bloomberg report published October 19 connects the founders of one of China's best-known open-weight AI labs to a research funder that UK intelligence has called a front for spying. Four Z.ai founders took part in a 2018 to 2021 research project co-funded by the China General Technology Research Institute (CGTRI), the organization MI5 said has "very strong ties" to China's main spy agency. Z.ai, also known as Zhipu AI, is the maker of the GLM models used in coding tools and cloud platforms around the world.

The story is easy to over-read, in both directions. MI5 did not name Z.ai, and nothing reported suggests the GLM models are compromised. But the funding link is documented, and it raises procurement and reputational questions. This post lays out what is claimed, what is verified, what is not known, and what it means for teams who use GLM. We worked from outlets citing Bloomberg, because the original is paywalled.

TL;DR: claimed vs verified

What Bloomberg reports

According to outlets summarizing the Bloomberg story, including Startup Fortune, the project ran from January 2018 to December 2021, which overlaps Z.ai's 2019 founding. It was co-financed by CGTRI and China's National Natural Science Foundation. Li Juanzi, a Tsinghua computer science professor and Z.ai co-founder, is the one participant named in the coverage we could read; we could not confirm the other three names. For the lab's leadership context, see our GLM 5.3 vision post.

One outlet frames the group as four of five founders; others say four founders. We use "four" because it is the common figure, and note that the exact company count is not something we could verify.

What MI5 actually said

The anchor for the story is MI5's first public "Espionage Alert", issued September 30, 2026, which named CGTRI. As reported by several outlets, the alert says:

• CGTRI has "very strong ties" to China's Ministry of State Security (MSS) and funds academic work that improves MSS espionage capability.
• More than 100 UK-based academics contributed to CGTRI-funded projects, in areas including AI, cybersecurity, covert communications and steganography.
• Many academics and institutions probably acted in good faith, and some may not have known who was funding the work.
• Universities should review current and planned collaborations immediately, and continued work could risk prosecution under the National Security Act 2023. Security Minister Dan Jarvis wrote to university vice-chancellors.

China's embassy in London called the allegations "imaginary and purely fabricated". We found no independent confirmation of MI5's intelligence assessment beyond the alert itself, which is typical for such warnings.

The good-faith language matters here. If UK academics could be unaware of the funder, the same caution applies to researchers in China, and the report does not establish what the Z.ai founders knew.

What is not established

1. Knowledge and intent. No source shows the founders knew CGTRI's alleged MSS links, or that the project served intelligence work.
2. A link to the models. The research ended in 2021. The GLM models were released years later. Reporting does not claim the history affects model integrity, data handling or behavior.
3. Any government action against Z.ai. We found no sanctions, bans or procurement orders aimed at the company over this. Contrast that with DeepSeek, where US civilian agencies were ordered to remove its technology, as in our DHS DeepSeek post.
4. Z.ai's side. We found no statement from the company or founders. Silence is not an admission.

Timeline

The gap between the project's end and the MI5 alert is nearly five years. That matters for two reasons. The alert is about CGTRI's current role, and the project predates the GLM generations that people use today. Whether CGTRI looked the same in 2018 as it does in MI5's 2026 assessment is a question the reporting does not answer.

What people are asking

Is Z.ai on a US or UK blacklist? We found no report that it is, and Bloomberg's story does not say so. DeepSeek has faced concrete US actions; Z.ai has not, as far as we could find.

Could academics at Tsinghua have known? The reporting describes a project co-financed by a national science foundation, which would look like ordinary academic funding from the inside. MI5 itself says many UK academics acted in good faith, which is the same argument in the other direction.

Why would a social-media analysis project matter? Mapping relationships between people from posts is a standard network-analysis task with both commercial and surveillance uses. The description alone does not show intent, but it is the sort of capability intelligence agencies buy, which is why the aim is quoted.

Does it change anything for open-weight users? The weights are public and can be audited, and nothing reported identifies a flaw. What changes is the reputational and procurement calculation for organizations that must document their supply chain.

Why it still matters

Even without technical findings, reputational and legal exposure is real for enterprises. Funding by an entity a Western government calls a spy front is the kind of fact that procurement teams, auditors and sanctions screeners notice. Z.ai's models are widely deployed: through cloud marketplaces such as Amazon Bedrock, in privacy products such as Brave's encrypted assistant, and as a teacher model in CoreWeave's distillation preview. Each of those operators may now be asked whether they have assessed this.

The lab is also under scrutiny for capability, not just funding. Anthropic raised concerns about open-weight cyber capability in its GLM 5.3 warning, and a legal-AI evaluation of GLM models drew its own claimed-versus-verified questions. The new report adds a governance thread to a lab that is already prominent in the open-weight ecosystem.

What to do if you use GLM

• Separate technical risk from provenance risk. Open weights can be inspected, run offline and sandboxed. A funding history does not change the weights you downloaded.
• Prefer self-hosting or a trusted cloud for sensitive data, and avoid sending regulated data to any vendor API you have not vetted. This applies to every provider, US or Chinese.
• Check your compliance rules. Government contractors, financial firms and universities in the UK may have their own screening requirements after the MI5 alert.
• Pin versions and hashes for model files, and verify checksums against the official Hugging Face release.
• Watch for primary sources. Look for a Z.ai statement, any follow-up from Bloomberg, new MI5 or UK government guidance, and any US action.

Our broader coverage of China's AI sector, including the US sanctions on MiraclePlus and Zhisuan, shows how quickly provenance questions can turn into policy. The pattern so far has been that governments act on technology transfer and model theft, not on researchers' funding histories, but that can change.

What would change our read

We would treat this as materially worse if documents show the project's outputs were delivered to security services, if a government names Z.ai in an advisory, or if independent analysis finds a technical issue in the weights or hosted APIs. We would treat it as largely resolved if Z.ai publishes a detailed account of the project, who funded it, and what the founders knew. Until one of those happens, the accurate summary is: a documented funding link, a serious alert about the funder, and no evidence connecting the two to Z.ai's products.

Related reading

• Z.ai GLM 5.3 coding benchmark release
• GLM 5.3 vision and community input: Tang Jie
• Anthropic's GLM 5.3 cyber capabilities warning
• GLM 5.3 on Amazon Bedrock
• DHS orders DeepSeek removed from federal civilian systems
• US sanctions MiraclePlus and Zhisuan
• China countermeasures after US model-theft accusation

Details are accurate as of October 20, 2026. The Bloomberg report is paywalled; we relied on outlets citing it and on reports of MI5's alert, and will update if Z.ai responds.