Open-Weight Model GLM-5.2 Said to Rival Anthropic’s Mythos in Cybersecurity
Chinese AI company Zhipu AI – also known as Z.ai – has released a new AI model, GLM-5.2, which according to security researchers comes close to Anthropic’s powerful Mythos model in certain cybersecurity scenarios. The model is available under open weights, meaning it can be freely downloaded, run locally, and modified. The move could shift the global balance of power in the AI race – and raises security-policy questions.
We’ve already reported on just how capable GLM-5.2 is. At its recent launch, benchmarks and tests showed that the Zhipu model even outperforms Google’s current top models.
What the researchers report
According to security experts, GLM-5.2 can match the latest US models when it comes to detecting software vulnerabilities and security flaws. On more general tasks, the reports say, the model still lags behind Anthropic’s and OpenAI’s products. Overall, however, the findings suggest that China has significantly narrowed the gap to the leading US systems – especially in cybersecurity.
Lior Div, CEO of the security firm 7AI, summed up the development to the Wall Street Journal by saying that China is steadily closing the gap.
The open-weights factor
Unlike the largely closed-off frontier models from the US, GLM-5.2 is an open model: anyone can run it on standard hardware and adapt it without oversight. This gives technically skilled users great flexibility and deep access. At the same time, it means that malicious actors could also deploy the model unobserved – a point observers highlight as a risk.
Security-policy context
The development comes at a tense moment in US policy. Washington has restricted China’s access to powerful models such as Anthropic’s Mythos and Fable, as well as to the hardware needed to run them; just this month, the US government ordered Anthropic to halt exports of a less capable variant on national-security grounds. Models capable of identifying vulnerabilities are viewed by the US government as a potential security risk.
Against this backdrop, the rise of GLM-5.2 fuels concern that restrictions on US frontier technology could end up handing Beijing an advantage in cyberwarfare – even as China catches up.
Possible consequences
The dual nature of the technology: The same capability that lets AI detect security flaws can be used for both defense and attacks. Researchers point out that it is now all the more urgent to deploy such models to close vulnerabilities before attackers can exploit them – otherwise a scenario described as “bugmageddon” could loom.
Intensified competition: Several Chinese providers are stepping up the pace. The security firm 360 Security Technology recently unveiled two AI-powered tools under the “Yitian Tulong” banner: Tulongfeng for automatically detecting vulnerabilities, and Yitianzhen for cyber defense and incident response. The company described Tulongfeng as China’s equivalent of Anthropic’s Mythos.
Cost pressure and adoption: Adoption of Chinese AI models is rising as companies look to cut costs. Microsoft and other providers are reportedly already considering hosting Chinese models – a development that could further shift the balance of power in the tech industry.
