Chinese Open-Weight AI Models Help Cybersecurity Firm Top HackerOne Hacking Contest

Chinese artificial intelligence models are showing strong capabilities in cybersecurity, after a US-based security company used open-weight models from Chinese AI developer Z.ai to finish at the top of a HackerOne vulnerability-finding competition.

Cybersecurity company Tenzai said its autonomous AI hacker ranked first on HackerOne's vulnerability disclosure programme for the second consecutive quarter.

This time, Tenzai used GLM 5.2 and GLM 5.3, open-weight models developed by Chinese AI company Z.ai. The models were used as part of Tenzai's automated system to search for security vulnerabilities in real-world software and infrastructure.

AI finds vulnerabilities in real-world systems

HackerOne is a platform where organisations work with security researchers to identify and report vulnerabilities in their systems.

Tenzai gave its AI hacker access to the platform and allowed it to search for security weaknesses.

According to Tenzai, the system was able to identify vulnerabilities while maintaining a false-positive rate of about 3.5%.

The company also said the latest run cost roughly 30% less per report than its previous run using US frontier AI models.

Tenzai's earlier tests showed that autonomous AI systems can identify complex vulnerabilities and combine multiple weaknesses into attack chains. The company has previously reported findings ranging from software vulnerabilities to remote-code-execution and database-access issues.

Open-weight models are becoming a cybersecurity issue

The result is significant because the models used by Tenzai are open-weight.

Unlike many commercial AI systems, open-weight models can be downloaded and run by developers and organisations without relying entirely on the model provider's online service.

This makes them easier to customise and deploy for different purposes.

It also means that the same technology can potentially be used by both defenders and attackers.

Tenzai CEO Pavel Gurvich has argued that open-weight models from China are already capable of performing sophisticated offensive cybersecurity tasks. The company said the models are not far behind leading US systems in cybersecurity-related capabilities.

The model is only part of the system

Tenzai said the HackerOne result does not necessarily mean that one AI model is simply better than another.

The company's system combines an AI model with an agentic security framework, which allows the AI to search for vulnerabilities, test potential weaknesses and produce reports.

Tenzai's CEO said the result demonstrated the importance of the system or "harness" surrounding the model, rather than just the underlying AI model itself.

This distinction is important because cybersecurity performance depends on more than a model's ability to understand code. The tools it can access, how it plans attacks, how it validates findings and how it interacts with target systems can all affect the final result.

Why open AI models matter to cybersecurity

The development creates a new challenge for security teams.

Powerful AI models that can be downloaded and operated outside a controlled commercial environment could allow organisations to build their own cybersecurity tools.

At the same time, the same capabilities could potentially be adapted for malicious purposes.

Recent research and industry testing have shown that increasingly capable AI agents can discover vulnerabilities, write code and automate parts of cyberattacks. Tenzai itself has reported that its AI hacker can continuously test applications and chain multiple vulnerabilities together.

This means defenders may need to consider AI not only as a security tool, but also as part of the changing threat landscape.

The race is moving beyond model development

The HackerOne results also show that the competition between AI systems is no longer limited to benchmarks that measure writing, reasoning or coding.

Cybersecurity is becoming another practical test of how well AI can operate in real environments.

For defenders, the development has two sides. AI can help security teams test more systems, identify vulnerabilities faster and reduce the cost of security testing.

But if similar tools become available to attackers, the speed at which vulnerabilities can be discovered and exploited could also increase.

As open-weight AI models become more capable and widely available, organisations may need to strengthen continuous vulnerability testing, access controls and monitoring to ensure that security weaknesses are discovered by defenders before attackers find them.