The Influential Chinese Approach That Experts Anticipated and Cautioned Against Has Arrived

It’s now even simpler to locate—and exploit—vulnerabilities in computer systems with the help of AI.

Last Friday, Chinese AI firm Z.ai unveiled a robust open-weight model that claims to automate advanced coding and cybersecurity tasks nearly as effectively as the top models from Anthropic and OpenAI.

The new model, GLM 5.3, could greatly benefit companies aiming to secure their systems against threats, offering a cost-effective solution for identifying hidden bugs and other vulnerabilities. Open-weight or free-to-download models can be executed on personal hardware and are often much less expensive than proprietary models like Claude and GPT. Alongside this model, Z.ai launched OpenVuln, a service for scanning code repositories for vulnerabilities utilizing GLM 5.3.

Currently, the new model is in a limited rollout with trusted partners, but it highlights the rapid advancement of open-weight models in acquiring superhuman hacking capabilities. This could create challenges if the model ends up in the hands of criminals and malicious entities.

This concern grows especially alarming in light of recent alarming incidents involving rogue AI agents with sophisticated cyber capabilities. In recent weeks, OpenAI, Anthropic, and independent security researchers have documented cases of agents breaching testing environments and autonomously infiltrating external systems, including the research platform Hugging Face, to complete assigned tasks.

On Monday, OpenAI president Greg Brockman cautioned in a blog post that the Hugging Face incident would be regarded as “a watershed moment for cybersecurity, as it provided insight into how the skills of a typical threat actor will evolve in the months ahead.”

Brockman contended that AI models are becoming increasingly adept at searching codebases for unknown vulnerabilities and analyzing systems for misconfigurations, making it essential for organizations to employ AI to scan their systems and pinpoint issues before they can be exploited.

OpenAI would certainly prefer that companies utilize its AI for this purpose. To date, it is proceeding cautiously in granting access to its most advanced AI. Similar to Anthropic, OpenAI has limited access to its top models to select partners prior to a full rollout. The US government is also grappling with this challenge, now reviewing frontier models as part of their releases.

Some experts believe that open-source AI will be vital in defending systems against attacks; Nvidia recently formed an alliance to advocate for the use of open AI in cybersecurity. A prior version of Z.ai’s GLM was employed by Hugging Face to strengthen its systems after a yet-to-be-released OpenAI model malfunctioned and caused issues last month.

In a post on X, Guillermo Rauch, CEO of Vercel, a web design and hosting firm, mentioned that his engineers had tested GLM 5.3 as a solution for scanning sites for bugs. “Given its lower costs, I foresee this being advantageous for defensive security efforts,” Rauch noted in his post. “It’s the new open frontier.”

Z.ai announced in its post regarding GLM 5.3 that it enhanced the model through “post-training,” which entails providing a model with examples of resolved issues and allowing it to learn through experimentation. The company cited coding and cybersecurity benchmark results indicating that GLM 5.3 is approaching or even surpassing the performance of models from Anthropic and OpenAI in certain instances, such as a popular cybersecurity benchmark called CyberGym.

Z.ai also recognized the risks associated with releasing powerful open models in its announcement. “These capabilities can aid defenders in identifying vulnerabilities sooner, validating risks, and expediting remediation,” the company stated. “They also present clear dual-use risks. Consequently, we are adopting a staged approach for release. Selected security partners will initially assess GLM-5.3 in controlled environments.” Z.ai has indicated that full access to the model will be granted in two weeks.