# China's Zhipu Claims Parity With Western Models on Cyber Defence and Publishes Its Own Standard

GLM-5.3 scored 84.5 percent on a code-vulnerability test against 83.8 percent for its American benchmark, while state media pressed developers to drop English artificial-intelligence terms.

- Published: 2026-08-18T05:17:18.667Z
- Canonical: https://polylog.news/2026-08-18/china-s-zhipu-claims-parity-with-western-models-on-cyber-def
- Publisher: Polylog (Global desk)
- Section: tech
- Sources: [South China Morning Post](https://www.scmp.com/tech/article/3364356/zhipu-ais-answer-project-glasswing-marks-shift-chinese-cyber-safety-researcher?utm_source=rss_feed), [South China Morning Post](https://www.scmp.com/news/china/politics/article/3364320/why-chinese-state-media-telling-people-stop-using-english-ai-terms?utm_source=rss_feed)

Zhipu AI, the Beijing model developer known internationally as Z.ai, released its GLM-5.3 model on August 14 alongside a programme it calls Shield of Open Source, which offers free security audits and automated code-auditing tools through its ZCode platform. A researcher told the South China Morning Post that the launch [marks a shift in how China's leading laboratories approach global cybersecurity](https://www.scmp.com/tech/article/3364356/zhipu-ais-answer-project-glasswing-marks-shift-chinese-cyber-safety-researcher?utm_source=rss_feed), describing it as the country's first answer to the Western Project Glasswing effort.

The published numbers show partial parity. Zhipu reported that GLM-5.3 scored 84.5 percent on CyberGym, a test of whether a model can read code, find security flaws and confirm that they are genuine, slightly ahead of the 83.8 percent it recorded for its American benchmark. On ExploitBench, which measures how far a model can carry an exploit, [the Chinese model scored 54.4 percent](https://www.scmp.com/tech/big-tech/article/3364077/zhipu-launches-flagship-model-glm-53-china-seeks-mythos-level-edge-cyber-defence) against 78 percent and 76.5 percent for the two leading American systems. These are the developer's own figures and have not been independently replicated. Z.ai said it will publish open weights about two weeks after launch, once safety evaluation and hardening are complete.

The technical push runs alongside a linguistic one. Chinese state media have called for English-language artificial intelligence terms to be replaced with standardised Chinese equivalents, a campaign the South China Morning Post described as an effort to build [what officials call discourse power](https://www.scmp.com/news/china/politics/article/3364320/why-chinese-state-media-telling-people-stop-using-english-ai-terms?utm_source=rss_feed). Terminology is where technical standards begin, and whoever writes the vocabulary tends to write the specifications that follow.

## What this means

Export controls were meant to hold Chinese models a generation behind, and a domestically trained system claiming parity on a defensive security benchmark weakens the argument that controls buy time. Chinese cloud and chip suppliers gain domestic demand as customers standardise on local models, while American vendors lose the assumption that their tools remain the reference point in Asian and Global South procurement. The terminology campaign matters because standards bodies follow vocabulary, and a separate vocabulary makes interoperability a political decision rather than a technical default.

## What to watch

- Whether independent researchers reproduce Zhipu's benchmark scores, since the entire parity claim rests on the developer's own testing.
- Whether the open weights are released on schedule, because a delay would indicate the company found capabilities it does not want in public hands.
- Adoption of Chinese-language AI terminology in national standards documents, which would show the discourse campaign moving from media into regulation.