China‑based AI lab Z.ai has warned that its latest model, GLM‑5.3, is so capable at finding and exploiting security flaws that the company will delay public release of the model weights for two weeks, according to a report by Axios. The lab said it is testing and strengthening safety and security controls before wider availability. The warning from Z.ai echoes concerns raised by Sam Altman of OpenAI and Dario Amodei of Anthropic, who have previously cautioned about the risks posed by powerful open‑weight models. Both executives have argued that models capable of cyber exploitation could be misused if released without strong safeguards, highlighting the growing tension between innovation and security in AI development.AI models capable of identifying and exploiting cyber vulnerabilities carry major implications for both sides of the security equation, offering real gains for defenders trying to patch systems, but also for hackers looking to scale up attacks.
What Z.ai announced
Recenlty, Z.ai warned that its newest model, GLM-5.3, has become so capable at finding and exploiting security flaws that the company is delaying the public release of the model’s weights by two weeks while it tests and strengthens its safety and security controls. The company specifically trained GLM-5.3 to improve at vulnerability discovery by letting it practice cyber tasks in controlled environments, and is rolling out a tiered access program that, for now, limits access to GLM-5.3 to selected security partners operating in controlled settings.
How it stacks up against US models
On CyberGym, a benchmark that measures how well models can find known security vulnerabilities, GLM-5.3 scored 84.5%, outperforming both Anthropic’s Fable 5 and OpenAI’s GPT-5.6 Sol. On ExploitBench, which evaluates a model’s ability to reason through and construct exploits for real vulnerabilities, GLM-5.3 trailed only Fable 5 and GPT-5.6 Sol among the models Z.ai tested it against.
Z.ai’s defenders-first framing
According to a new disclosure site the company released recently, Z.ai says its GLM models have already identified more than 2,400 security flaws, including over 1,000 rated as critical or high severity, with some of those vulnerabilities found in the Linux kernel and in widely used VMware and Apache projects. The company is framing GLM-5.3’s release primarily as a benefit to defenders rather than a risk, writing on X that an open world cannot have only open attack surfaces, arguing that it must also have an open shield. Alongside the model announcement, Z.ai also introduced a new program letting open-source maintainers have a GLM model scan their repositories for bugs.Z.ai’s models have already shown some practical value on the defensive side: Hugging Face has said it used GLM-5.2 to investigate a recent breach involving OpenAI’s models, after guardrails built into U.S. frontier models declined to assist with that investigation.
The catch
Z.ai acknowledged in its X post that once GLM-5.3’s weights become public, the company will have no way to control how people modify or use the model going forward, a limitation inherent to open-weight releases more broadly. Meanwhile, the Trump administration appears to be weighing potential approaches to regulating open-source AI models as their capabilities increasingly rival those of their closed, proprietary counterparts.