Introducing GLM-5.3: Equipped with Enhanced Cyber Features and Uncovered a Significant Vulnerability in Cursor

The Chinese AI company Z.ai, recognized globally for its expanding range of effective open-source GLM series language models, has unveiled its latest version, GLM-5.3. This release boasts significant enhancements in long-term coding capabilities and notable advancements in cybersecurity features that could raise concerns due to their sensitivity.

At the outset, GLM-5.3’s cybersecurity skills have reportedly identified a “potentially serious vulnerability in Cursor,” an AI coding startup recently acquired by SpaceX. This information was shared by Z.ai developer advocate Lou on the platform X, with VentureBeat also mentioning the same for verification.

Initially, GLM-5.3 can only be accessed through Z.ai’s GLM Coding Plan and ZCode coding environment. API access and open weights will follow later after thorough safety evaluations, as stated by the company. Z.ai expects to release the weights roughly two weeks after the initial launch.

For enterprise developers, the key highlight is not merely another set of performance metrics. Z.ai clarifies that GLM-5.3 uses the same foundational model as GLM-5.2; improvements stem exclusively from scaling post-training across diverse environments, various tasks, and additional reinforcement-learning computations. This approach tests the limits of how much can be accomplished with a base model without incurring the high costs of retraining.

“All we did for GLM-5.3 was scale post-training,” Z.ai noted in a technical release.

The results indicate substantial growth potential but have also given rise to an unexpected challenge for open-model developers: Z.ai found that the cybersecurity capabilities advanced more quickly than anticipated as training progressed, particularly in moving from recognizing vulnerabilities to creating full exploitation pathways.

Recently, Z.ai announced it would implement controls around some of the model’s advanced features, including a “trusted access” method for sensitive functionalities.

Significant advancements in coding without a new foundation model

GLM-5.3 builds on the 743-billion-parameter model that powered GLM-5.2, expanding the post-training system originally designed for long-term reinforcement learning.

These environments increasingly mimic complete engineering tasks instead of just isolated programming problems.

Z.ai presents scenarios where the model has access to codebases, documentation, compute clusters, storage systems, and experimental outcomes, tasked with diagnosing issues, modifying systems, conducting experiments, and showing measurable improvements while maintaining accuracy. Some tasks aim to reflect several days of work for a skilled engineer.

This strategy has led to significant enhancements in Z.ai’s reported evaluations from one generation to the next.

GLM-5.3’s scores on Terminal-Bench 3.0 increased from 4.6 to 28.3, from 46.2 to 66.9 on DeepSWE v1.1, and from 26.2 to 48.2 on AutomationBench. On Agents’ Last Exam CLI, it went up from 23.8 to 28.5.

However, it does not surpass every competitor in the field. For instance, Z.ai’s benchmark comparisons show GPT-5.6 Sol scoring 34.6 and Claude Fable 5 at 33.7 on Terminal-Bench 3.0, against GLM-5.3’s score of 28.3. On DeepSWE v1.1, GLM-5.3 achieved 66.9, while GPT-5.6 Sol scored 72.7 and Fable 5 reached 69.7.

Z.ai emphasizes operational efficiency over purely benchmarking success.

In its private Z.ai Code Bench, GLM-5.3 achieved a 34.5% result at its maximum reasoning setting while utilizing approximately 75,000 output tokens per task. In comparison, GLM-5.2 scored 23.4% using around 96,000 tokens. At a high-effort level, GLM-5.3 recorded 31.4% using roughly 50,000 tokens, whereas Z.ai reported 29.5% for Claude Opus 4.8, which consumed 120,000 tokens.

Since Code Bench results are proprietary, they should be viewed as in-house reported outcomes rather than independent assessments. Nevertheless, reducing token use while enhancing task completion is crucial for businesses deploying coding agents, as extensive loops can accelerate inference costs and delays.

Rapid enhancements in cybersecurity capabilities

Cybersecurity has seen particularly unexpected advancements.

Z.ai integrated vulnerability-discovery environments into GLM-5.3’s post-training process with the hope of improving the model’s ability to identify software flaws. Contrary to expectations, the company noted that the capabilities began progressing toward constructing comprehensive exploitation chains.

“Cyber capability progressed faster than anticipated as we scaled post-training,” Z.ai commented.

On CyberGym, which evaluates vulnerability discovery against source code, GLM-5.3 achieved 84.5%, compared to GLM-5.2’s 77.2%. This score also edges out the reported scores for GPT-5.6 Sol at 83.6% and Mythos 5 at 83.8%.

However, GLM-5.3’s advantage does not cover the entire exploitation spectrum. It scored 54.4% on ExploitBench, which is a significant improvement from GLM-5.2’s 24.4%, but still trails behind the 76.5% reported for GPT-5.6 Sol and 78% for Mythos 5.

On ExploitGym, GLM-5.3 completed 105 tasks within a normalized two-hour budget and 130 tasks under six hours, up from 29 and 39 for GLM-5.2, while Fable 5 completed 181 and 247 and GPT-5.6 Sol reached 216 and 293.

The direction of progress may be more critical than ranking positions.

Z.ai stated that collaboration with security teams in China led to the discovery of 2,436 vulnerabilities across 269 projects, following expert evaluation, screening, and deduplication. Its disclosure ledger counts 1,097 as critical or high severity, with 53 publicly revealed and 2,383 still under embargo as of the release date.

This situation highlights a growing challenge for leading model providers: the same sophisticated long-horizon capabilities that enhance models for software development can also elevate them as security researchers—and possibly as offensive entities.

Developers need to adapt how they call the model

Developers transitioning existing GLM applications should note a major change in API behavior.

GLM-5.3 allows three reasoning-effort levels—low, high, and max, with max being the default and recommended for coding tasks. However, unlike previous versions, the thinking function cannot be disabled.

Applications currently set to thinking.type: "disabled" must change this to enabled and specify a reasoning effort prior to changing the model identifier to GLM-5.3. Otherwise, the request will fail, making this an actual migration rather than a straightforward swap of model names.

Z.ai’s rapid evolution toward automated engineering

GLM-5.3 signifies a key milestone in Z.ai’s swift shift—previously known as Zhipu AI—toward automated coding agents and extensive autonomous engineering tasks.

GLM-4.5, released in July 2025, set much of this precedent. The 355-billion-parameter mixture-of-experts model aimed to integrate reasoning, coding, and agent functionalities, while its smaller counterpart, GLM-4.5-Air, consisted of 106 billion total parameters. Z.ai launched these models with open weights focused on compatibility with agent frameworks.

Following in September, GLM-4.6 enhanced context length from 128,000 to 200,000 tokens, targeting coding, tool usage, and agent workflows spanning various environments, including Claude Code, Cline, Roo Code, and Kilo Code. Z.ai shifted its focus to improve token efficiency in real-world coding assessments instead of benchmark performance solely.

A more considerable architectural transition occurred with GLM-5 in February 2026, expanding from GLM-4.5’s 355 billion parameters to 744 billion with 40 billion active parameters and raising pretraining data to 28.5 trillion tokens. It introduced an asynchronous reinforcement-learning infrastructure and explicitly realigned the GLM series around “agentic engineering” and prolonged tasks.

By June, GLM-5.2 offered a stable 1-million-token context window, open weights under an MIT license, and support in over 20 coding environments. The model also introduced IndexShare, aiming to alleviate significantly the computational demands for extensive contexts.

GLM-5.2 was priced at $1.40 per million API input tokens and $4.40 per million output tokens, with reduced costs for cached inputs, presenting Z.ai as both a technical and cost competitor against proprietary rivals.

In addition to model development, Z.ai’s ambitions have extended significantly. Recently, reports indicated that the company raised approximately HK$31.4 billion, nearly $4 billion, through a share sale in Hong Kong, with the proceeds intended for R&D, computing infrastructure, talent acquisition, and business expansion.

Collectively, these releases reflect a consistent evolution: GLM-4.5 combined reasoning, coding, and agents; GLM-5 dramatically scaled the base model; GLM-5.2 pivoted towards long-context and long-horizon engineering; and GLM-5.3 is now focusing on maximizing capabilities from the existing foundation through post-training.

Pricing, ZCode, and availability

GLM-5.3 is now available through Z.ai’s GLM Coding Plan and ZCode.

ZCode serves as the company’s coding-agent environment, designed to manage long-running “Goal” tasks that involve planning, implementation, testing, and verification of work. This tool also allows remote control of ongoing tasks and is compatible with macOS, Windows, and Linux.

Current pricing for individual GLM Coding Plans begins at a promotional rate of $12.60 per month for Lite, offering 10,000 credits weekly. The Pro option is priced at $56 monthly, providing six times the Lite usage, while the Max plan costs $117.60 per month, enabling 14 times the Lite usage. Team Standard and Premium pricing is set at $88 and $188 per user per month, respectively.

Z.ai has transitioned the Coding Plan to a points-based quota system that separately tracks input, cached input, and output tokens. Calls made outside the company’s weekday peak hours consume half the standard points.

As for general GLM-5.3 API pricing, the company has not released specific figures, making it difficult to directly compare production costs with GLM-5.2 or other competing frontier models until staged API access becomes available.

This gradual release may prove to be the most significant aspect of GLM-5.3.

Z.ai has dedicated the past year to promoting an open-model strategy focused on permissive weights, low-cost inferences, and integration with existing coding-agent ecosystems. GLM-5.3 illustrates the outcomes when this approach excels, particularly in a sensitive sector: improved autonomous engineering equates to enhanced autonomous security research.

Consequently, this model propels Z.ai’s coding ambitions while also compelling the company to grapple with the complex trade-offs between capability and access that larger closed frontier labs face.

For enterprise developers, GLM-5.3 represents two notable aspects worth monitoring. Its coding results suggest that increasingly capable agents can arise from enhanced post-training and diverse environments without the continual need to rebuild foundational models. Additionally, its cybersecurity results underscore why determining the method of agent distribution may become just as crucial as deciding on their training.