Z.ai recently announced via X that its GLM model successfully identified 23 vulnerabilities within the Windows operating system, all of which have been confirmed by Microsoft’s Security Response Center (MSRC). This achievement is not merely a byproduct of AI testing in security; it serves as a testament to the model’s capabilities.
The company emphasized that its findings extend beyond open-source projects, marking a significant advancement in commercial software security through independent vulnerability discovery and responsible disclosure. The emphasis on “independent” is crucial, as it underscores the model’s ability to operate autonomously in identifying security flaws.
This accomplishment is the latest in a series of successful disclosures by Z.ai, which has meticulously built a public record over the latter half of 2026. The journey began with open-source code, and the Windows vulnerabilities represent a new chapter in a narrative that has gained increasing attention. In August, Z.ai released GLM-5.3, achieving an impressive score of 84.5% on CyberGym, a benchmark for assessing a model’s proficiency in detecting and exploiting vulnerabilities. This score surpassed competitors like Anthropic’s Mythos 5 and OpenAI’s GPT-5.6 Sol, which scored 83.8% and 83.6%, respectively. Shortly after this launch, GLM-5.3 flagged a significant vulnerability in Cursor, an AI-powered code editor, demonstrating its practical application in real-world software.
Since the release of GLM-5.2, Z.ai’s models have collectively uncovered over 4,200 vulnerabilities across approximately 390 open-source projects, with more than a thousand categorized as critical to high severity. Rather than simply announcing these figures, Z.ai has established a public Security Disclosure Ledger to meticulously track each finding through a coordinated disclosure process. This approach ensures that technical details remain confidential with verified maintainers until a fix is implemented, reflecting the diligent and methodical nature of a genuine security vendor rather than a mere marketing initiative.
The acknowledgment of 23 vulnerabilities by MSRC, while modest compared to the extensive open-source findings in Z.ai’s ledger, carries a different weight. Open-source maintainers often operate with limited resources and may be more inclined to credit an AI tool without extensive scrutiny. In contrast, Microsoft’s MSRC is recognized for its rigorous vulnerability triage operations, making the confirmation of these findings a significant milestone for Z.ai.
Why the commercial pivot matters
Z.ai has framed the Windows findings as evidence that GLM-powered discovery has progressed beyond the realm of open-source projects. The challenge posed by commercial software is notably greater, as Windows lacks a comprehensive GitHub issue tracker for models to reference. Successfully identifying genuine flaws and obtaining confirmation from Microsoft’s security team serves as a clearer indicator of the model’s capabilities than merely adding to an open-source tally, regardless of its size.
This distinction holds importance beyond mere accolades. Reports indicate that NIST’s AI safety division has flagged GLM-5.3 due to concerns about its cyber capabilities, particularly since Z.ai releases its model weights openly. This openness allows anyone to download and modify the model, potentially removing its safety features. Anthropic has noted that a modified version of GLM-5.3 with diminished safety protocols exhibited a significant decrease in its refusal rate for harmful requests. In response to this scrutiny, Z.ai has chosen not to retreat into silence but to increase its output of disclosures, acknowledgments, and evidence demonstrating that the same capabilities raising regulatory concerns are also effectively identifying and reporting genuine vulnerabilities through MSRC and coordinated disclosure.
The 23 acknowledgments from MSRC represent 23 valuable data points, and it is anticipated that Z.ai will continue to target commercial software in its ongoing efforts.
Also read:
- UK CFOs now say AI will shrink their own workforce, Deloitte finds
- China’s open AI models now trail the US frontier by just seven months
- China’s Humanoid Robot Shipments Hit 97% of the World as Tesla Stalls
This article is posted in AI News, check it out for more related stories.
