Chinese AI company Z.ai has said its latest model, GLM-5.3, has approached Anthropic’s Mythos 5 in some cybersecurity tests, although the results have not been independently verified.
Z.ai said GLM-5.3 scored 84.5 per cent on CyberGym, a benchmark that evaluates an AI model’s ability to inspect code, identify vulnerabilities and verify whether the flaws are genuine. The company reported a score of 83.8pc for Mythos 5.
However, GLM-5.3 performed significantly worse than Mythos 5 when it came to turning identified vulnerabilities into functioning exploits, an important capability in defensive cybersecurity research.
On the ExploitBench test, Z.ai said its model scored 54.4pc, compared with 78pc for Mythos 5. In separate timed assessments, GLM-5.3 completed 105 attack-development tasks in two hours and 130 in six hours, while Mythos 5 completed 181 and 247 tasks respectively.
Anthropic has made Mythos, a cybersecurity-focused version of its Claude Fable 5 model with certain safeguards removed, available only to vetted organisations. The restrictions reflect concerns that AI capable of identifying and exploiting software vulnerabilities could benefit defenders while also making cyberattacks easier to conduct.
Z.ai said it planned to publicly release GLM-5.3 in about two weeks after completing security assessments and improving its safeguards. Its most sensitive cybersecurity capabilities will be restricted to verified users through a “trusted access” programme, the company said.
In a post on X on Friday, Z.ai said initial access would be provided to a select group of launch partners before being expanded through what it described as a “consistent and responsible process”.
AI governance researcher Gabriel Wagner of Concordia AI said the move appeared to be an early example of a Chinese AI company publicly citing safety concerns as a reason for delaying the open release of model weights.
Z.ai said it had introduced multiple safeguards into GLM-5.3, including systems designed to screen potentially dangerous requests, monitor the model’s activities and train it to reject malicious tasks.
The company said the measures were intended to distinguish harmful activities from legitimate uses such as software debugging, cybersecurity education and authorised security testing.
Critics, however, have raised concerns that such safeguards could be harder to maintain once an AI model is made available for download, modification or integration with external tools.
Z.ai has presented GLM-5.3 as an alternative to restricted-access cybersecurity models such as Mythos, arguing that advanced defensive tools should also be available to open-source developers and smaller security teams.
The company said it would launch an “Open Source Shield” initiative to audit selected open-source projects, provide model access for defensive purposes and integrate code-auditing capabilities into its ZCode programming product.
Mr Wagner described the initiative as resembling Anthropic’s limited-access Project Glasswing, but with greater emphasis on openness.
US-based AI company Hugging Face said last month that it had used Z.ai’s earlier GLM-5.2 model to help defend against a cyberattack by a rogue OpenAI agent that had gained access to its systems.
Z.ai is not the first Chinese company to claim capabilities comparable to Mythos. Cybersecurity firm 360 said in June that its Tulongfeng vulnerability-discovery system had achieved similar capabilities by combining AI models with security data and automated tools. Those claims, however, have also not been independently verified.
GLM-5.3 differs from purpose-built cybersecurity systems in that it is a general-purpose coding model that Z.ai says developed its cyber capabilities through additional post-training and reinforcement learning.
The company said GLM-5.3 uses the same base model as GLM-5.2 but was trained across longer and more diverse task environments.
The launch builds on growing international interest in GLM-5.2, which has gained attention among overseas developers for coding and AI-agent capabilities that users and analysts said were approaching those of leading US models while being available at a lower cost.
