Chinese AI developer Zhipu has introduced GLM-5.3, a coding-focused model that exhibits impressive cybersecurity skills, positioning it closely to some of the top global competitors in vulnerability detection. Despite its advances, it still lags behind these peers in more complex exploitation tasks.
According to Zhipu's testing metrics, GLM-5.3 slightly surpasses Anthropic's Mythos 5 and OpenAI's GPT-5.6 Sol on the CyberGym benchmark for vulnerability identification and validation, achieving a score of 84.5%. The competitors scored slightly lower, with Mythos 5 at 83.8% and GPT-5.6 Sol at 83.6%. However, the model falls substantially behind on the ExploitBench, recording just 54.4%, compared to Mythos 5's 78% and GPT-5.6 Sol's 76.5%.
“GLM-5.3 stands out as the most capable open-weights model for coding, boasting a 50% improvement over its predecessor, GLM-5.2, in our proprietary Z.ai Code Bench,” Zhipu stated. They noted that the latest model not only identifies vulnerabilities but also formulates comprehensive strategies for complete exploitation.
The firm highlights the advancements made over GLM-5.2, particularly in its ExploitBench score, which jumped from 24.4% to the current figure. In practical tests on ExploitGym, it executed 105 exploitation scenarios within two hours and 130 in six hours, a considerable increase from the 29 and 39 tasks completed by GLM-5.2 during the same time frames.
Zhipu attributes these enhancements to a combination of post-training strategies and reinforcement learning amidst increasingly complex task simulations.
Neil Shah, VP for research at Counterpoint Research, noted the implications of the evolving capabilities of coding models. He stated, “We are witnessing a moment where training an AI to excel as a software engineer inadvertently equips it with hacking skills.” He emphasized that the logical processes required for debugging and testing code overlap significantly with those used by cyber attackers to exploit weaknesses.
Identification of Vulnerabilities
Zhipu has collaborated with security teams in China to assess their models against real-world coding environments. The findings indicate that the GLM-5.3 model identified a total of 2,436 vulnerabilities across 269 projects, including 1,097 categorized as medium-to-high severity.
The vulnerabilities span a range of critical systems including device kernels, operating systems, browser engines, and various web applications, as well as network protocols. According to Zhipu, its security disclosure ledger has cataloged 107 critical and 990 high-severity vulnerabilities. Out of these, 53 have been made public while 2,383 remain under wraps. Notably, the oldest vulnerability identified dates back to 1981, with many lingering in the code for an average of 26.6 years before detection.
However, Zhipu has not disclosed the proportion of these findings that were previously unknown or whether they have been confirmed by independent sources. The model’s findings are tracked through the Z.ai Security Disclosure Ledger as they navigate the disclosure protocol.
Shah characterized these developments as a double-edged sword for security teams. “AI tools can enhance system audits and bug fixes significantly,” he explained. “Yet, once the weights of an AI model are released to the public, any safety measures could be compromised without oversight.”
Scaling the Same Base Model
Interestingly, Zhipu believes the advancements of GLM-5.3 arise not from a new foundational model, but from scaling post-training initiatives. The company enriched its training environments to simulate more authentic and prolonged work settings. One exercise required the model to diagnose computational bottlenecks, implement optimizations, conduct experiments, and deliver tangible improvements.
The company also integrated vulnerability-discovery data into the training regimen, which contributed to reported gains not only on its internal metrics but also on public coding benchmarks.
This growing interconnectedness between programming and offensive security strategies complicates traditional categorizations, according to Shah. He reiterated, “The patterns an AI employs to troubleshoot and optimize code mirror the tactics attackers use to locate vulnerabilities and launch exploits.”
Implications of Open-Weight Release
Zhipu plans to release the weights for GLM-5.3 approximately two weeks following the launch, contingent on their safety assessment processes. This pending release is aimed at making the advanced model available, which has demonstrated capabilities that stretch from vulnerability detection to more refined exploitation techniques.
However, details on additional safeguards accompanying this release remain unspecified. Shah raised concerns regarding how quickly vulnerabilities could transition from identification to exploitation once such capabilities are publicly accessible. He asserted, “If these AI tools can uncover numerous flaws in operational systems and are available for anyone to use, the response time to potential attacks diminishes drastically.”
As cybersecurity challenges escalate, defensive strategies must evolve concurrently, necessitating that control mechanisms be integrated into the development and deployment stages of AI models and autonomous systems.
The original article appeared on InfoWorld.
