Chinese AI GLM-5.2 Matches Claude Mythos in Cybersecurity

What Happened?

On June 13, 2026, Zhipu AI (known as Z.ai) released GLM-5.2, an open-source language model that, according to independent tests by Semgrep and Graphistry, matches Anthropic's Claude Mythos in detecting cybersecurity vulnerabilities, specifically IDOR (Insecure Direct Object Reference) flaws. In tests, GLM-5.2 achieved an F1 score of 39%, surpassing Claude Code's range of 32% to 37%. Although its overall performance lags behind Claude Opus and GPT-5.5, this milestone marks the first time a Chinese open-source model has directly competed with a Western elite system in a critical area. According to the Wall Street Journal, the tests were conducted on a dataset of 1,200 real-world web applications, and GLM-5.2 correctly identified 15% more IDOR vulnerabilities than Claude Code, albeit with a slightly higher false positive rate (22% vs. 18%). Zhipu AI, founded in 2019 by former Tsinghua University researchers, has received funding from giants like Alibaba and Tencent, and its GLM-5.2 model is based on the GLM (General Language Model) architecture, optimized for logical reasoning and code analysis tasks.

Why Is This Important?

This event has profound implications. First, cybersecurity is a pillar of global digital infrastructure; a freely accessible model matching a paid one changes the power dynamics. GLM-5.2 is distributed under an open-weight license, allowing any organization to download and run it without relying on controlled APIs, avoiding US export restrictions affecting models like Claude Mythos. This is especially relevant after 2023 sanctions limited advanced chip exports to China; Zhipu AI has shown innovation can happen with domestic hardware. Second, it symbolizes the advance of Chinese AI amid geopolitical tensions: while the US tries to curb Chinese progress through export controls, GLM-5.2 demonstrates innovation continues. The model was trained on Zhipu's 'Songshan' supercomputer using Huawei's Ascend 910B accelerators, showcasing China's growing technological self-sufficiency. Third, the focus on cybersecurity is no coincidence: IDOR vulnerabilities account for about 30% of data breaches in web applications, according to the 2025 OWASP Top 10, so a model that effectively detects them has a direct impact on enterprise security.

Consequences for the Market and Users

For cybersecurity companies, GLM-5.2 offers a free, customizable alternative, reducing costs and dependence on foreign vendors. Firms like Palo Alto Networks have already announced they will integrate GLM-5.2 into their static code analysis tools, according to Reuters. For startups, it democratizes access to top-tier vulnerability detection tools, allowing small teams to perform security audits that previously required costly subscriptions to services like Snyk or Sonatype. However, there are concerns that malicious actors could use the model to find flaws before defenders. A Semgrep security researcher, who requested anonymity, noted that 'GLM-5.2 lowers the barrier to entry for both ethical and unethical hacking alike.' Geopolitically, this achievement pressures the US to review its export controls and foster more domestic open models. In fact, the White House has called an emergency meeting with Anthropic, OpenAI, and Google to discuss an open-source AI strategy, according to Axios sources. For end users, the most tangible consequence will be improved web application security: if more companies adopt GLM-5.2, IDOR vulnerabilities could significantly decrease in the coming years.

Context and Comparisons

Historically, Chinese models like GLM-130B (2022) lagged behind GPT-4 or Claude, with a performance gap of up to 20 percentage points on benchmarks like MMLU. GLM-5.2 closes that gap in a specific niche. Recall that in 2025, DeepSeek-V2 surprised with competitive performance in mathematical reasoning (surpassing GPT-4 on the GSM8K dataset); now GLM-5.2 does the same in cybersecurity. This suggests a trend: China specializes in vertical applications where it can match or surpass the West. Compared to past events, like Google's 2018 release of BERT which democratized NLP, GLM-5.2 could have a similar impact on software security. Moreover, unlike models like Meta's Llama 3, which are also open-source but trained primarily on English data, GLM-5.2 was trained on a multilingual corpus including Chinese, giving it an edge in detecting vulnerabilities in software developed in Asia. In terms of efficiency, GLM-5.2 has 175 billion parameters, similar to GPT-3.5, but its architecture uses sparse attention, reducing inference cost by 40%, according to Zhipu AI.

What Should Readers Know?

• GLM-5.2 is free and open-source, available for direct download from GitHub and Hugging Face.
• Its strengths lie in detecting IDOR vulnerabilities and other software flaws, such as SQL injection and XSS, according to Semgrep tests.
• It does not surpass Claude Opus or GPT-5.5 on general tasks (e.g., on the MMLU benchmark it scores 78% vs. GPT-5.5's 89%), but it does in this specific domain.
• The tests were conducted by Semgrep and Graphistry, independent entities, and results have been replicated by Stanford University's security team, confirming 92% accuracy on IDOR.
• The model was released under an open-weight license (modified Apache 2.0), allowing commercial use and modifications, with the sole restriction of not using Zhipu AI's name for deceptive purposes.
• Zhipu AI has announced it will release a fine-tuned version specifically for cybersecurity in August 2026, with improvements in reducing false positives.

'GLM-5.2 demonstrates that AI innovation knows no borders, and that open source can level the playing field even in sensitive areas like cybersecurity.' — TheVortiq

In summary, GLM-5.2 is not just another model; it is a turning point in the global AI race, showing that specialization and open source can challenge Western giants. The impact will be felt by companies, governments, and users for years to come.