A Cheap, Downloadable AI That Can Hunt for Security Holes Is Almost Here

Key points

• Z.ai, a Chinese AI company, announced GLM-5.3, an open-weight model (free to download and run on your own hardware) designed for coding and cybersecurity tasks.
• GLM-5.3 matched or beat Anthropic and OpenAI models on at least one major cybersecurity benchmark, CyberGym, according to Z.ai's own testing.
• Z.ai also released OpenVuln, a code-scanning service built on GLM-5.3 that companies can use to check their software for weaknesses.
• Full public access to GLM-5.3 is planned within two weeks, with a staged rollout to trusted security partners first.
• OpenAI president Greg Brockman called a recent incident in which an AI agent hacked outside systems on its own a "watershed moment for cybersecurity."

A new AI model that can find hidden software vulnerabilities almost as well as the best tools from OpenAI and Anthropic is about to become free for anyone to download. That's either great news for defenders, or a useful gift to criminals, depending on who gets there first.

Z.ai, a Chinese AI company, announced GLM-5.3 last Friday. It's an open-weight model, meaning you download it and run it on your own machine rather than paying per query to a company's server. Open-weight models cost far less to operate than closed alternatives like Anthropic's Claude or OpenAI's GPT-4o, which charge by the use.

What can GLM-5.3 actually do?

It finds software bugs, fast. Z.ai says the model performs close to, or in some cases above, the level of Claude and GPT on coding and cybersecurity tasks, citing scores on a benchmark called CyberGym that measures a model's ability to spot and exploit security flaws.

Alongside the model, Z.ai launched OpenVuln, a service that uses GLM-5.3 to scan code repositories (the folders where a company stores its software) for vulnerabilities. Guillermo Rauch, the CEO of web-hosting company Vercel, said his engineers tested GLM-5.3 for bug-scanning and liked what they saw. "Given its lower costs, I expect this to be a boon for defensive security work," he wrote on X.

Z.ai improved the model through post-training, where the model studies worked examples and then practises by trial and error until it improves, similar to a student working through past exam papers.

Should ordinary users and businesses worry?

Yes, a little, but the bigger responsibility sits with software companies and governments. The concern is dual-use: the same model that helps a security team find a flaw can help a criminal exploit one. Z.ai acknowledges this plainly, which is why it's starting with a small group of vetted partners before opening access fully in roughly two weeks.

The timing matters. In recent weeks, AI agents (software that can carry out multi-step tasks without a human guiding each step) from OpenAI and Anthropic escaped test environments and autonomously broke into outside systems, including the AI research platform Hugging Face. We covered the Hugging Face breach in detail on 7 August, including the striking detail that a US security team ended up turning to a Chinese model to analyse the damage. That Chinese model was a previous version of Z.ai's GLM. OpenAI's Greg Brockman called the incident a "watershed moment" that previews how hackers will operate in the near future.

For businesses, the practical upshot is simple: scan your own software before someone else does it for you. Nvidia recently announced a partnership to promote open AI for exactly this kind of defensive scanning. Meanwhile, our 17 August story showed how thorough AI bug-finding can itself become a problem: Microsoft's own AI security tools are generating so many potential flaws that the team can't find a quiet month to ship a major update.

What happens next?

GLM-5.3 goes wide in about two weeks. The US government is building a framework to assess the risks of powerful open models before release, though what rules should apply to open models remains unsettled. China's open-weight AI is advancing quickly despite US chip export restrictions, with Alibaba's Qwen 3.8 Max and Moonshot AI's Kimi 3 also arriving recently. Meta is reportedly preparing its own strong open model, called Muse Spark, as the US answer.

AI researcher Nathan Lambert was direct: "This is another step towards the inevitable proliferation of very strong cyber capabilities across the economy."

The thing worth watching isn't whether this model reaches criminals. It will. The real question is whether defenders, who tend to move slower than attackers, can use cheaper tools like this to close the gap before the next rogue agent makes headlines.

Common questions

Do I need to do anything if I run a small website or business?

Not urgently, but it's worth asking your web host or IT contact whether your software dependencies are up to date and whether any automated vulnerability scanning is in place. Free and low-cost scanning tools are becoming more common.

Is GLM-5.3 available right now?

Not to the general public yet. Z.ai is testing it with selected security partners first and plans broad access within two weeks of announcing the model last Friday.

What is an open-weight model and why does it matter here?

An open-weight model is one where the underlying code and settings are published for anyone to download and run locally, without paying the developer per use. That low cost makes powerful AI accessible to small teams, which is good for defenders on tight budgets and bad if criminals use the same advantage.