Zhipu's GLM-5.3: Accelerated Cybersecurity Skills Challenge Competitors

Chinese AI developer Zhipu has rolled out its latest coding model, GLM-5.3, which exhibits unexpectedly advanced skills in cybersecurity. This new model reportedly approaches the levels of top global models in identifying vulnerabilities, although it still lags behind in executing deeper security exploits.

In Zhipu’s evaluations, GLM-5.3 scored 84.5% on CyberGym, a benchmark assessing vulnerability discovery. This positions it slightly ahead of Anthropic's Mythos 5, which scored 83.8%, and OpenAI's GPT-5.6 Sol with 83.6%. However, the performance gap widens on ExploitBench, where GLM-5.3 achieved 54.4%, while its competitors scored 78% and 76.5%, respectively.

Zhipu claims that GLM-5.3 is now the leading open-weights coding model, boasting a 50% improvement compared to its predecessor, GLM-5.2, in internal benchmarks like Z.ai Code Bench. The firm noted that as they scaled post-training, the cyber capabilities of the model advanced beyond expectations, with GLM-5.3 evolving from simple vulnerability identification to generating coherent exploitation strategies.

The model's ExploitBench score skyrocketed from a mere 24.4% with GLM-5.2, while it completed significantly more exploitation tasks—105 tasks within two hours and 130 within six hours, compared to 29 and 39 by its predecessor. This leap in performance is largely credited to advancements made during post-training, where reinforcement learning was leveraged across progressively more complex scenarios.

Neil Shah, VP of research at Counterpoint Research, pointed out that as coding models gain sophistication, they may also inherently develop offensive capabilities. "Teaching an AI to be an accomplished software engineer may lead to it also becoming an effective hacker," he remarked. He emphasized that the reasoning applied by coding AIs for debugging is the same as what attackers use to identify and exploit weaknesses.

Significant Vulnerabilities Identified

Zhipu is actively collaborating with security teams in China to evaluate its model against actual codebases. The results of these assessments are telling: the model identified 2,436 vulnerabilities across 269 projects, with a total of 1,097 categorized as medium-to-high severity.

These findings encompass a wide range of components, from operating systems and browser engines to open-source infrastructures and web applications. Notably, Zhipu's security disclosure ledger lists 107 critical and 990 high-severity issues, with 53 findings already made public while 2,383 remain embargoed. The oldest vulnerability in this dataset dates back to 1981, highlighting a concerning average discovery time of 26.6 years for vulnerabilities.

Shah described the dual nature of this capability for security teams. While AI tools can expedite system audits and bug-fixes, their release into the public domain without any protective measures could enable malicious actors to exploit them. "Once an AI model's weights are made freely available, safety mechanisms can be stripped away, making it harder to control misuse," he added.

Enhanced Pre-Training Model

Zhipu's inference on GLM-5.3's capabilities centers on post-training enhancements rather than the creation of an entirely new base model. By expanding their training environments to reflect real-world scenarios—such as diagnosing bottlenecks and implementing optimizations—the model's application of learned skills improved significantly.

The company’s methodology now includes vulnerability-discovery data within the training phases, which has resulted in a reported 50% performance enhancement over the previous model, GLM-5.2, and improvements on external coding and agent benchmarks.

Open-Weight Release Considerations

Zhipu plans to make GLM-5.3's model weights public in approximately two weeks, pending safety evaluations and hardening procedures. The open-weight model is expected to demonstrate not just vulnerability discovery capabilities but also increasingly nuanced exploit reasoning.

However, Zhipu has not detailed what additional safeguards will accompany this release. Shah raised a significant concern regarding the rapid potential for the exploitation of vulnerabilities: “If anyone can download AI tools capable of discovering unpatched flaws, the reaction time to these threats diminishes drastically.” He advocated for built-in defenses in the development and deployment processes of such sophisticated AI models.

This article originally appeared on InfoWorld.