Introduction: AI Is Changing the Cybersecurity Battle
Artificial intelligence is making cybersecurity more powerful, but it is also creating new risks.
On October 29, 2026, a new report about China’s Z.AI and its GLM-5.3 model highlighted how quickly AI systems are advancing in complex cybersecurity tasks. According to a report by Reuters, the model completed at least one attempt in three of 13 multi-stage cyberattack simulations conducted by AI security firm Irregular.
The tests showed that the model could perform several steps involved in simulated cyberattacks, including accessing protected services, creating alternative network connections, and coordinating activity across multiple machines.
These findings do not mean that GLM-5.3 has independently carried out confirmed attacks against real-world targets. However, they highlight an important concern: AI models are becoming more capable of performing technical tasks that previously required considerable human expertise.
For businesses, developers, and website owners, the message is clear. AI can help security teams identify weaknesses faster, but organizations must also prepare for attackers who use similar technology.
In this article, we explain what the GLM-5.3 cybersecurity tests revealed, why open-weight AI models matter, and how businesses can strengthen their security in 2026.
What Is GLM-5.3 and Why Is It Making Headlines?
GLM-5.3 is an AI model developed by Z.AI, formerly known as Zhipu AI, a Chinese artificial intelligence company.
Unlike a basic chatbot designed primarily to answer questions, advanced AI models can assist with multi-step technical tasks. Depending on the tools, permissions, and environment provided, these systems may analyze code, identify software weaknesses, execute commands, and work through complex problems.
GLM-5.3 has attracted attention because of its cybersecurity capabilities and its availability as an open-weight model.
Open-weight models make their trained model weights available for developers to download and use under the applicable license. This can give organizations more flexibility over how they deploy AI systems.
However, open weights do not automatically mean that a model is completely open source, free of restrictions, or safe for every use.
The same flexibility that benefits legitimate developers can also complicate security controls. Once a model is available for independent deployment, its original developer may have less direct control over how every user operates it.
This is one reason why the GLM-5.3 findings matter beyond a single AI company.
What Did the GLM-5.3 Cybersecurity Tests Reveal?
According to Reuters’ October 29 report, Irregular tested GLM-5.3 in 13 complex, multi-stage cyberattack scenarios.
The model completed at least one attempt in three scenarios. Irregular described this as its strongest result for an openly available model in the relevant tests.
The assessment also reported that the model could access protected services, establish alternative network connections, and coordinate operations across multiple machines in simulated environments.
These capabilities are important because complex cyberattacks rarely depend on one action. They can involve several connected stages, from identifying a weakness to navigating a network and attempting to access protected information.
An AI system that can assist with multiple stages may reduce the time and technical knowledge required to perform certain activities.
However, the results need context.
The assessment involved controlled simulations, and completing an attempt in three scenarios does not mean the model achieved every objective or successfully breached three real organizations.
The tests also cannot establish that every deployment of GLM-5.3 behaves in the same way. Results can change depending on available tools, instructions, safeguards, and the testing environment.
How does GLM-5.3 compare with leading AI models?
The US National Institute of Standards and Technology (NIST), through its Center for AI Standards and Innovation (CAISI), published an assessment of GLM-5.3 on September 17, 2026.
According to NIST’s evaluation, GLM-5.3 was the most cyber-capable open-weight model released at the time of its assessment.
However, it remained behind the leading US models evaluated by the organization, with an aggregate capability gap estimated at approximately four months.
NIST also tested the model across several cybersecurity benchmarks, including vulnerability discovery and exploit development.
The results demonstrate that GLM-5.3 represents meaningful progress in open-weight AI cybersecurity, but they do not show that it has surpassed all competing frontier models.
It is also important to remember that different organizations use different benchmarks and testing conditions. Results from separate evaluations should not be treated as a direct, like-for-like ranking.
Why Open-Weight AI Models Create New Security Challenges
Open-weight AI models offer several advantages. Developers can run them on supported infrastructure, customize their applications, and potentially reduce dependence on a single AI provider.
Businesses may use these systems for code analysis, internal security testing, vulnerability research, and other technical work.
However, open-weight models also introduce challenges.
1. More people can access advanced capabilities
When model weights are publicly available, more developers and organizations can experiment with the system.
This can support innovation and improve access to cybersecurity tools. It can also make it harder for the original developer to prevent every possible misuse after release.
2. Safety controls may differ between deployments
AI companies can implement safeguards in hosted services, but developers who run downloadable models may configure their systems differently.
Some deployments may include strong monitoring and access controls. Others may have limited safeguards or none beyond the model’s own behavior.
As a result, the security of an AI application depends on more than the model itself.
3. AI can reduce the technical barrier to cybercrime
Some cyber activities require specialized knowledge, significant time, or familiarity with complex software systems.
AI can help users understand code, automate repetitive work, and investigate technical problems. In the wrong hands, these capabilities may also assist malicious activity.
This does not mean that AI automatically makes every attack successful. Real-world systems contain additional technical barriers, and AI-generated attempts may fail.
Nevertheless, organizations should prepare for the possibility that attackers will use AI to work more efficiently.
AI-Powered Cyberattacks: What Businesses Should Watch For
The GLM-5.3 findings are part of a wider shift in cybersecurity.
AI tools can support both defensive and offensive activities, and security professionals are increasingly concerned about the speed at which software vulnerabilities can be identified and investigated.
Here are four areas businesses should monitor.
Automated vulnerability discovery
AI systems can analyze source code and help identify potential weaknesses. This can help developers find bugs before attackers exploit them.
However, the same capability may help malicious actors identify weaknesses in applications that have not been updated.
Businesses should maintain an accurate inventory of their software, monitor security advisories, and prioritize patches for vulnerabilities affecting exposed systems.
AI-assisted phishing
AI can help generate convincing messages, adapt writing styles, and produce content in different languages.
Employees may therefore encounter fraudulent messages that are more polished and personalized than traditional phishing attempts.
Organizations should use email filtering, multifactor authentication, employee training, and verification procedures for sensitive requests.
Automated attack workflows
AI agents can work through multi-step tasks when given suitable tools and permissions.
In cybersecurity, this can help authorized teams investigate incidents. It can also increase risk if an AI agent receives excessive access or is allowed to interact with systems outside its authorized scope.
Organizations should limit permissions, isolate testing environments, and monitor agent activity.
Misuse of stolen credentials
AI cannot eliminate the need for valid access, but it may help an attacker use available information more efficiently.
Businesses should protect credentials, rotate exposed secrets, remove unnecessary permissions, and monitor unusual account activity.
These measures remain important regardless of which AI model an attacker might use.
Can AI Also Make Cybersecurity Stronger?
Yes. AI is not only a cybersecurity risk. It can also help organizations defend their systems.
Security teams can use AI to review code, summarize security alerts, investigate suspicious activity, and prioritize vulnerabilities.
For example, a security analyst dealing with thousands of alerts may use an AI-assisted tool to group related events and identify those requiring immediate investigation. A developer may use AI to understand a software vulnerability and prepare a proposed fix for human review.
These applications can improve efficiency, especially for smaller organizations with limited security resources.
However, AI-generated findings should be validated. A model can produce incorrect conclusions, overlook important context, or recommend a fix that creates another problem.
AI should support security professionals rather than replace essential controls and expert judgment.
For a broader explanation of this shift, read our guide: AI Is Changing Cybersecurity: How Artificial Intelligence Is Fighting Cyber Threats.
The key lesson is that defenders need to adopt useful AI capabilities while maintaining strong security practices.
What Website Owners Should Do to Reduce AI-Related Cyber Risks
Small businesses and website owners should not assume that advanced AI-powered attacks only affect large technology companies.
Websites can face risks from outdated software, compromised administrator accounts, vulnerable plugins, exposed credentials, and poorly configured servers.
For WordPress website owners, several practical steps can make a meaningful difference.
Keep WordPress, themes, and plugins updated
Install security updates promptly, particularly when a vulnerability affects a publicly accessible component.
Remove unused plugins and themes instead of leaving them installed.
Enable multifactor authentication
Protect administrator accounts with multifactor authentication wherever possible. Use unique passwords and a password manager to reduce the risk of credential reuse.
Maintain reliable backups
Keep regular backups of website files and databases. Store at least one backup separately from the main hosting environment and test the restoration process.
A backup is useful only if it can be restored successfully.
Use a web application firewall
A reputable web application firewall can help filter malicious requests and block certain common attack patterns.
It is not a replacement for patching, secure configuration, or monitoring, but it can provide another layer of defense.
Monitor website activity
Review administrator logins, unexpected file changes, suspicious redirects, and unusual traffic patterns.
Website owners should also check that contact forms, email subscriptions, and other important functions work correctly after security changes.
Prepare an incident response plan
Know how to disable compromised accounts, restore a clean backup, contact your hosting provider, and investigate suspicious activity.
These steps can reduce the impact of a security incident even when the precise attack method is unknown.
The Growing Connection Between AI Models and Computing Infrastructure
The development of powerful AI models also depends on access to computing infrastructure.
Training and running advanced AI systems require specialized processors, data centers, energy, and networking. This is why competition among AI developers increasingly overlaps with competition among chipmakers and cloud providers.
Google’s Tensor Processing Units (TPUs), for example, are part of the wider competition to supply the computing resources needed for AI development.
For more context on this industry trend, read our report: OpenAI Plans to Buy Google AI Chips: What It Means for Nvidia and the AI Industry.
Although computing infrastructure and cybersecurity are different subjects, they are connected by the rapid expansion of AI capabilities. More accessible computing and more capable models can create opportunities for innovation while increasing the importance of responsible deployment.
Readers can also explore our coverage of broader industry developments in Top 5 AI News Today: Claude Security Incidents, OpenAI and Google Cloud Partnership, and AI Industry Changes.
What Should Happen Next?
The GLM-5.3 findings highlight the need for more rigorous AI security testing.
AI developers should evaluate models before release, study how safeguards perform under realistic conditions, and clearly communicate important limitations.
Organizations deploying AI agents should use restricted permissions, isolated environments, activity logging, and clear rules about which systems agents are authorized to access.
Independent researchers and public institutions also have an important role. Transparent evaluations can help businesses understand what AI models can do, where they fail, and what protections are needed.
At the same time, security assessments should avoid exaggerating results. A model completing a simulated attack is evidence of a capability under specific conditions, not proof that it has independently compromised real-world targets.
The goal should be to understand the risks accurately and develop defenses before those risks become widespread.
Conclusion: AI Cybersecurity Risks Are Becoming Harder to Ignore
The GLM-5.3 cybersecurity findings are another sign that AI is changing how security work is performed.
Open-weight models can expand access to advanced technology, help developers improve software, and support security research. They can also create new challenges when powerful capabilities are used without sufficient safeguards.
For businesses, the most practical response is to strengthen existing defenses, keep software updated, secure credentials, monitor systems, and use AI tools responsibly.
For AI developers, the priority should be better evaluations, safer deployment practices, and clear boundaries for automated agents.
The future of cybersecurity will not depend on AI alone. It will depend on how effectively people, organizations, and technology work together to prevent attacks and respond when something goes wrong.
Frequently Asked Questions (FAQs)
1. What is GLM-5.3 AI?
GLM-5.3 is an AI model developed by Z.AI, formerly known as Zhipu AI. It has attracted attention for its reasoning, coding, and cybersecurity capabilities, including performance in controlled vulnerability and cyberattack simulations.
2. Did GLM-5.3 carry out real-world cyberattacks?
The October 29, 2026, report from Reuters described tests conducted in controlled simulations. Those findings do not establish that GLM-5.3 independently carried out confirmed cyberattacks against real-world organizations.
3. Why are open-weight AI models a cybersecurity concern?
Open-weight models allow developers to access and operate model weights independently. This flexibility supports research and innovation, but it can also make it harder for the original developer to control how every deployment is used.
4. Can AI be used to prevent cyberattacks?
Yes. AI can help security teams analyze code, identify suspicious activity, investigate alerts, and prioritize vulnerabilities. However, its findings should be validated, and AI tools should complement established security controls.
5. How can businesses protect themselves against AI-powered cyberattacks?
Businesses should update software, enable multifactor authentication, protect credentials, monitor suspicious activity, maintain tested backups, use appropriate security filtering, and prepare an incident response plan.
6. Does AI make cyberattacks easier?
AI can reduce the time and technical expertise required for some cybersecurity tasks. However, it does not guarantee that attacks will succeed. Results depend on the target, the model, the tools available, and the security controls in place.
7. What is the difference between open-weight AI and open-source AI?
Open-weight AI makes trained model parameters available under specified terms. Open-source AI generally involves broader openness, potentially including source code, training information, and other components. The exact meaning depends on the license and materials released.
8. Why is AI cybersecurity important in 2026?
AI systems are becoming more capable of assisting with complex technical work. Organizations need to understand both the defensive benefits and the potential misuse of these systems as they develop security policies and update their defenses.
Editorial note: This article reflects reporting and assessments available as of October 29, 2026. Cybersecurity benchmark results depend on the evaluation method and testing environment; they should not be interpreted as proof of successful real world attacks.
External sources
