China's Zhipu AI (Z.ai) GLM -5.2 is one of the most important milestones in the global artificial intelligence race: an open-weight model that, according to independent benchmarks, equals or even surpasses Anthropic 's Claude Mythos in identifying critical software vulnerabilities. GLM-5.2 was released on June 13, 2026 and has already caused strong reactions in both the technology industry and US national security circles.
See also: Claude Fable 5: Anthropic releases first public Mythos-Class model
Zhipu AI, a company with roots in Tsinghua University in Beijing, developed GLM-5.2 using a Mixture-of-Experts (MoE) with 744 billion parameters. Notably, the model was trained exclusively on domestic Huawei Ascend chips, demonstrating China’s ability to circumvent US restrictions on semiconductor exports. The model also has a context window of 1 million tokens, allowing it to process entire code bases in real time — a capability that was previously the exclusive preserve of leading proprietary models.
Researchers at Semgrep found that GLM-5.2 achieved an F1 score of 39% in detecting IDOR (Insecure Direct Object Reference) vulnerabilities , compared to Claude Code ’s 32% — and at a cost of only ~$0.17 per inference . This finding surprised even the researchers themselves, as an open-weight Chinese model managed to outperform one of the most advanced proprietary systems in the West in such a specialized and critical domain.
Technical characteristics and cybersecurity capabilities
In addition to its performance in detecting IDOR bugs, GLM-5.2 stands out for a number of technical features that make it particularly attractive to cybersecurity professionals. The output speed reaches 400 tokens per second in the high-speed variant, while the maximum output capacity is 131,072 tokens per instance. The model natively supports English and Chinese, with strong performance in more than 15 languages. Particularly important is that it can run on a Mac with 256 GB of memory via Unsloth, eliminating the need for data center infrastructure.
GLΜ -5.2 is released under the MIT license, which allows for free local installation and use by anyone. This feature makes it extremely flexible for developers and security researchers, but at the same time raises serious concerns: without central control, malicious users can exploit it for offensive security operations, vulnerability detection in unauthorized targets, or even malware. Zhipu AI has also chosen to incorporate low levels of built-in moderation, which enhances flexibility but increases the risk of abuse.
See also: OpenAI Daybreak: The Answer to the Claude Mythos for Cybersecurity
GLM-5.2 and geopolitical implications: The threat to national security
The Trump administration has labeled advanced AI models capable of detecting vulnerabilities — such as Anthropic ’s Mythos and Fable — as serious national security threats, and has worked hard to restrict China’s access to them. However, the release of GLM-5.2 demonstrates that these efforts have not prevented Chinese companies from developing equivalent capabilities. Zhipu AI has already been on the US Entity List since January 2025 , which restricts its dealings with American companies but does not prevent the distribution of open-weight models.
It is worth noting that OpenAI recently released GPT-5.6, which has also raised concerns about potential abuse in the cybersecurity sector, resulting in access restrictions. Meanwhile, Chinese company 360 Security has developed the Tulongfeng, which according to the Wall Street Journal is also comparable to Mythos in some vulnerability detection scenarios. The global cybersecurity AI market is expected to reach $25 billion by 2027, with automated bug detection being a key growth segment.
Industry experts have expressed their admiration for GLM-5.2 's performance . Matt Velloso , a former vice president at Meta , DeepMind , and Microsoft , called it " the first open model to cross the line into an everyday tool ," while Jeremy Howard , co-founder of Fast AI , said it was " at least as good as Opus 4.8 and GPT 5.5 ." On the Arena AI platform , GLM-5.2 ranks first worldwide for frontend coding among available models.
See also: Anthropic's Mythos found vulnerabilities in US government systems
GLM -5.2 represents a significant paradigm shift in the field of artificial intelligence: an open-weight model that challenges the most expensive closed systems, runs on consumer hardware, and excels in critical cybersecurity applications. For Greek businesses and security professionals, its availability opens up new possibilities for cost-effective penetration testing and vulnerability detection — but at the same time requires increased vigilance, as the same capabilities are also accessible to malicious users.
