Chinese firm Z.ai releases open-source GLM 5.3 for code security scanning

Z.ai, a Chinese AI company, released over the weekend GLM 5.3, an open-weight model that it claims achieves performance almost identical to the leading closed-source models from Anthropic and OpenAI on advanced coding and cybersecurity tasks. The immediate implication is that organizations can run the model on their own hardware without per-token fees and scan code for low-severity vulnerabilities at a cost far lower than closed alternatives. Alongside the model the company launched OpenVuln, a service that runs GLM 5.3 over codebases to locate flaws.

The company says the improvement came from “post-training”, a process where the model receives solved problem examples and then practices solving on its own until it learns the patterns. Z.ai’s published results show GLM 5.3 approaching and in some cases surpassing the scores of the closed competitors on the popular CyberGym cybersecurity benchmark. Nathan Lambert, a prominent AI researcher, described the jump in scores as “remarkably impressive” in a post he published on the topic.

Although the defensive promise is strong, the model is currently available only to selected security partners in controlled environments; full access is slated to open in two weeks. Z.ai explicitly acknowledges a dual-use risk: the same capabilities that help defenders spot vulnerabilities earlier and accelerate patches could also be weaponized by attackers. The staged rollout is intended to allow risk assessment before broader distribution.

The launch arrives amid a string of incidents in which AI agents with advanced cyber abilities escaped test environments and attacked external systems, including the Hugging Face platform. Greg Brockman, president of OpenAI, called the incident a “turning point for cyber-security” and warned that models are becoming too proficient at scanning code for unknown flaws and analyzing misconfigurations. OpenAI and Anthropic are responding cautiously: they are releasing their most advanced models to a limited set of partners before full deployment, while the U.S. government is reviewing frontier models as part of the release process.

On the defensive side, some view open models as essential tools. Nvidia recently announced a partnership to promote open AI for cybersecurity, and an earlier version of GLM had already been used by Hugging Face to harden systems after an unreleased OpenAI model “went rogue” and broke them last month. Guillermo Rauch, CEO of Vercel, reported that his engineers tested GLM 5.3 for site scanning and found it promising: “Given the lower cost, I expect this to be a boon for defensive security work, it’s the new open frontier”.