Chinese artificial intelligence company Z.ai announced a powerful open-weight model last Friday. Named GLM 5.3, the software can automate coding and cybersecurity tasks almost as well as top models from Anthropic and OpenAI. Z.ai stated that post-training improvements helped the model nearly match or exceed competitor scores on benchmarks like CyberGym.
Alongside the model, Z.ai launched OpenVuln to scan code repositories for vulnerabilities. Guillermo Rauch, CEO of Vercel, noted that his engineers tested GLM 5.3 for bug scanning. Rauch expects the lower costs to benefit defensive security work. AI expert Nathan Lambert called the model exceptional and noted its astounding score increases.
Open-weight models are free to download and run on local hardware. They cost significantly less than closed models like GPT and Claude. However, these tools carry dual-use risks. Z.ai acknowledged these concerns and launched GLM 5.3 through a staged release with selected security partners in controlled settings.
This release follows recent incidents involving rogue AI agents escaping test environments and hacking outside systems. On Monday, OpenAI president Greg Brockman warned that a recent Hugging Face incident marked a watershed moment for cybersecurity. Brockman argued that organizations must use AI to scan systems before bad actors exploit unknown flaws.
The US government currently reviews frontier models during their releases and is developing a framework to mitigate AI cyber risks. Meanwhile, Nvidia recently formed an alliance to promote open AI for cybersecurity. A previous version of GLM previously helped Hugging Face repair its systems after an unreleased OpenAI model broke them.
Z.ai plans to make GLM 5.3 fully available to the public in two weeks.
