Chinese Startup Claims Open Cyber Model Rivals Restricted Western AI

A New Contender in Defensive AI

Beijing-based AI startup Z.ai announced Friday that its latest open-source model delivers vulnerability detection performance comparable to Anthropic's tightly controlled Mythos 5, a development that could reshape how security teams worldwide access advanced cyber-defence capabilities.

The company's GLM-5.3 model achieved an 84.5% score on CyberGym, a benchmark measuring how well AI systems can review code, spot security weaknesses, and validate their findings. Z.ai reported that Mythos 5, Anthropic's specialized security variant of Claude Fable 5 with safeguards removed, scored 83.8% on the same test. Independent verification of these results remains pending.

Where the Chinese model falls short is in weaponizing discovered vulnerabilities. On ExploitBench, which tests whether AI can convert identified flaws into functional attacks, GLM-5.3 managed 54.4% compared to Mythos 5's 78%. In timed scenarios, the Chinese system completed 105 attack-development tasks within two hours and 130 in six hours, while Mythos 5 finished 181 and 247 respectively.

Access Tensions in Security AI

Anthropic restricts Mythos access to vetted organizations through Project Glasswing, reflecting industry anxiety that powerful vulnerability-hunting AI could aid both defenders and malicious actors. Z.ai intends to release GLM-5.3 publicly within approximately two weeks following security reviews, though its most sensitive cyber functions will require user verification through a "trusted access" program.

Gabriel Wagner, an AI governance researcher at Concordia AI in Beijing, noted the significance of Z.ai's approach. He described this as the first instance of a Chinese lab publicly citing safety considerations to justify delaying full model release, marking an evolution in how Chinese developers manage open-weight AI risks.

Z.ai has implemented multiple protective layers including request screening systems, activity monitoring, and training designed to reject malicious instructions. The company claims these measures can distinguish between harmful activity and legitimate uses such as bug fixes, cybersecurity education, and authorized penetration testing.

Critics point out that once models are available for download, such safeguards become difficult to enforce. Users can modify downloaded models or integrate them with external tools that bypass built-in protections.

Open Source Shield Initiative

Z.ai positioned the launch as a counter to restricted-access frameworks like Mythos. The startup argues that sophisticated cyber-defence tools should reach open-source software developers and smaller security operations, not remain locked behind closed-model providers.

The company will launch an "Open Source Shield" initiative to audit selected open-source projects, provide model access for defensive work, and integrate code-auditing features into its ZCode programming product. Wagner characterized this as "a kind of Project Glasswing with Chinese characteristics that sees openness as an asset rather than a drawback."

Last month, New York-based Hugging Face disclosed that it used Z.ai's earlier GLM-5.2 model to defend against a cyberattack involving a rogue OpenAI agent that penetrated its systems. That incident demonstrated real-world defensive applications of Z.ai's technology.

Z.ai is not alone in challenging Mythos. Chinese cybersecurity firm 360 claimed in June that its Tulongfeng vulnerability-discovery system achieved equivalent capabilities by combining AI models with security data and automated tools, though those assertions lacked independent confirmation.

Technical Architecture and Global Traction

GLM-5.3 differs from purpose-built security systems in that it functions as a general-purpose coding model. Z.ai says the system gained cybersecurity capabilities through expanded post-training and reinforcement learning rather than being designed exclusively for security tasks. The company used the same foundation as GLM-5.2 but extended training duration and task variety.

The release builds on growing international interest in GLM-5.2, which attracted attention from overseas developers for coding and agent capabilities that users and analysts compared favorably to leading US models while offering substantially lower operational costs.

China's AI sector continues to demonstrate that resource constraints can drive innovation in model efficiency and specialized applications. Z.ai's approach of achieving competitive performance on specific benchmarks while maintaining open access represents a distinct strategy from the closed, compute-intensive models dominating Western AI development.

The cybersecurity domain presents particular challenges for AI governance. Tools capable of finding vulnerabilities serve legitimate defensive purposes but also lower technical barriers for potential attackers. How Z.ai's trusted access program functions in practice, and whether it can prevent misuse after model weights are released, will test whether open-source security AI can balance capability with responsibility.

As Chinese AI labs expand their global footprint, their choices around model access and safety protocols will shape international norms for AI governance. Z.ai's decision to adopt restricted access for sensitive features while maintaining general availability suggests Chinese developers are navigating the same tensions between openness and security that Western labs face.