Chinese artificial intelligence developer Zhipu has introduced GLM-5.3, an advanced coding-focused AI model that has shown notable capabilities in cybersecurity. According to Zhipu, GLM-5.3 has rapidly developed skills in identifying vulnerabilities, positioning itself alongside leading global models while still falling short in deeper exploitation tasks.
In testing, GLM-5.3 achieved an impressive 84.5% on CyberGym, a benchmark for vulnerability recognition. This performance slightly surpasses Anthropic’s Mythos 5, which scored 83.8%, and OpenAI’s GPT-5.6 Sol at 83.6%. However, when evaluated on the more complex ExploitBench, GLM-5.3’s score of 54.4% lagged significantly behind its competitors—Mythos 5 scored 78%, and GPT-5.6 Sol reached 76.5%.
“GLM-5.3 is the most capable open-weights model for coding, with a 50% enhancement over our previous GLM-5.2 on the in-house Z.ai Code Bench,” Zhipu remarked. The model has transitioned from merely pinpointing individual vulnerabilities to devising comprehensive strategies for entire exploitation chains.
GLM-5.3 also shows considerable improvement compared to Zhipu's prior version, GLM-5.2. Its ExploitBench score more than doubled, rising from 24.4% to 54.4%. On ExploitGym, the new model successfully completed 105 exploitation tasks in two hours and 130 in six hours, showcasing a significant jump from GLM-5.2's 29 and 39 tasks, respectively. Zhipu attributes these advancements to an enhanced post-training approach that includes reinforcement learning across progressively challenging environments.
Neil Shah, VP of research at Counterpoint Research, highlighted a growing concern as coding AI models gain proficiency in cybersecurity tasks. “When we train AI to excel as software engineers, it inadvertently learns the skills to become an effective hacker,” he explained. This dual-use nature of coding models raises critical implications for security teams, amplifying the significance of governance for these technologies.
Significant Vulnerabilities Uncovered
Zhipu has also collaborated with security teams in China to assess its AI against real-world coding scenarios. Their model reportedly identified 2,436 vulnerabilities across 269 projects, which include 1,097 issues rated at medium-to-high severity. The vulnerabilities span a wide range of targets, including system kernels, operating systems, and web applications.
The company maintains a security disclosure ledger that details its findings, which include 107 critical and 990 high-severity issues. While 53 vulnerabilities have been made public, the majority—2,383—remain under embargo. The vulnerabilities discovered date back as far as 1981, with many having lingered in codebases for an average of 26.6 years prior to detection.
Shah referred to this capability as a double-edged sword for security personnel. “AI systems can audit and rectify flaws in systems rapidly. However, once the AI model’s weights are made publicly accessible, the inherent safety measures can be easily bypassed,” he cautioned.
Post-Training Scaling Enhances Performance
Zhipu attributes the enhancements seen in GLM-5.3 to scaling during post-training rather than the development of a new base model. The firm expanded its training scenarios to simulate extended and realistic professional workflows. For instance, the model has been tasked with diagnosing bottlenecks and implementing optimizations while balancing correctness with demonstrable improvements.
By integrating vulnerability discovery data into its training framework, GLM-5.3 has shown a 50% improvement over GLM-5.2 on Z.ai Code Bench, as well as positive outcomes on publicly available coding benchmarks.
Shah emphasized the blurring lines between coding and offensive cybersecurity as these models evolve. “The reasoning that an AI employs to examine code and fix errors is the same method an attacker uses to exploit vulnerabilities,” he noted.
Releasing Open Weights Heightens Risks
Zhipu intends to release the model weights for GLM-5.3 approximately two weeks post-launch, following a comprehensive safety review. This decision reflects the growing acknowledgment of the capabilities now accessible in the AI model, ranging from identifying vulnerabilities to executing sophisticated exploitation strategies.
However, as the company has yet to clarify what additional safeguards will accompany the open-weight release, this lack of transparency raises concerns. Shah warned that the speed with which vulnerabilities could transition from identification to exploitation could shrink dramatically. “If AI tools are discovering numerous unaddressed flaws and those capabilities become publicly available, the time to respond diminishes almost to nothing,” he asserted.
This indicates that protecting against rapid machine-speed attacks necessitates implementing controls at the development and deployment stages of AI models and autonomous agents.
The article originally appeared on InfoWorld.
