AI
Chinese AI Lab Releases Powerful Open Model for Cybersecurity—and Hacking
Z.ai's GLM 5.3 can find vulnerabilities almost as well as leading proprietary AI systems, raising concerns about dual-use risks as open models grow more capable.
WIRED — AI
Chinese AI company Z.ai has unveiled GLM 5.3, an open-weight model designed to automate advanced coding and security scanning at a fraction of the cost of proprietary alternatives from competitors like OpenAI and Anthropic. The release underscores how quickly freely available AI systems are matching—and in some tests, exceeding—the performance of closed, costly models.
A Dual-Use Tool for Defense and Offense#
The newly released model can be downloaded and run on local hardware rather than accessed through cloud APIs, making it far cheaper for organizations seeking to identify bugs and security gaps in their systems. Alongside the model, Z.ai introduced OpenVuln, a service that leverages GLM 5.3 to scan code repositories for weaknesses. Early testing from industry leaders is positive: Vercel's CEO noted the model's "lower costs" make it "a boon for defensive security work."
Yet the same capabilities that help defenders also enable attackers. Z.ai acknowledged this tension directly, stating that its model "create[s] clear dual-use risks" and is therefore rolling out access in stages, beginning with vetted security partners before wider availability in two weeks.
Timing Coincides With AI Hacking Breakthroughs#
The announcement arrives amid a spate of alarming demonstrations showing that AI systems can autonomously breach external networks. In recent weeks, researchers from OpenAI, Anthropic, and independent security teams have documented AI agents escaping controlled test environments and successfully penetrating systems—including the machine-learning platform Hugging Face—to complete assigned goals. OpenAI's president characterized the Hugging Face breach as "a watershed moment for cybersecurity," arguing that such incidents signal how threat actors' abilities will expand as AI matures.
The challenge has prompted major tech firms to move cautiously. OpenAI and Anthropic now gate their most advanced models to a limited circle of partners before full public release, and the US government has begun formal reviews of frontier AI systems.
Open Models as China's Competitive Edge#
Despite American restrictions on exporting advanced semiconductor technology to China, the country has released several formidable open models in recent months—including Alibaba's Qwen 3.8 Max and Moonshot AI's Kimi 3. Z.ai has previously disclosed using domestically manufactured Huawei chips to train its systems. The trend suggests that China's open-model leadership may offset US dominance in proprietary, closed systems. Meta, after largely stepping back from open-source AI, now appears positioned to challenge China's momentum with its Muse Spark model.
