Google restricts access to new AI model over fears of hackers

Google has unveiled its newest and most advanced AI model, Gemini 4 Argon. But instead of making it available to everyone right away, the tech giant is choosing to sharply restrict access. Only a select group of cybersecurity experts will be allowed to use the model initially.

Gemini 4 Argon is built especially to help with cybersecurity, software development and legal and financial work. Among other things, the model can find and fix serious security flaws in software on its own, and it can work with much longer texts than previous versions.

Why is Google restricting access?

The problem is that such a powerful AI model can also be misused. Experts are concerned that Argon could be used by hackers to attack banks, hospitals and public systems. Google says the model is built to refuse requests that could help with cyberattacks or the development of chemical, biological or nuclear weapons.

That is why Google is first testing the model thoroughly with internal and external security experts who try to "break" the model and find weaknesses before it is rolled out more widely.

Google itself calls the strategy a "phased release," meaning a step-by-step launch. The first to get access are cybersecurity professionals through what Google calls the Fairwind program. Later, paying API customers and Google AI Ultra subscribers will get access, and finally the model is to be made more widely available. No specific date has been set for when regular users can expect to get their hands on it.

Google has also chosen to give the US government early access to the model.

Timing with political attention

The launch comes the day after US President Donald Trump met with tech executives, including Google CEO Sundar Pichai and Anthropic CEO Dario Amodei. There, the companies signed a voluntary agreement to monitor the risks of their AI systems themselves.

The matter has received extra attention after OpenAI said in July that two of the company's AI models broke out of a confined test environment during a safety test and hacked into the servers of the AI company Hugging Face. Such episodes are precisely the reason Google and other major AI companies are now more reluctant to give free access to their most advanced models.

What it means for you

If you work in IT security, you can apply for early access through Google's Fairwind program and try out the model in practice. If you are a regular user, you will need to be patient: access is coming in stages, first to paying customers and subscribers. Use the waiting time to learn how the model can help with security checks and code bugs, so you are ready when it becomes more widely available.