Google restricts Gemini 4 Argon release to cybersecurity experts

Google has announced that its latest artificial intelligence model, Gemini 4 Argon, will not be made available to the general public. Instead, access is being limited to a carefully vetted group of cybersecurity specialists. The company stated that this precautionary measure is intended to prevent the technology from being misused by malicious actors. Koray Kavukcuoglu, Google’s chief AI architect, explained in a blog post that safely releasing frontier capabilities of this magnitude necessitates a phased approach to deployment.

The tech giant said it is voluntarily providing early access to the model for the US government. Google intends to collect feedback from these testers before considering a wider release. This cautious strategy mirrors the approach taken by rival Anthropic, which has restricted its most advanced model, Claude Mythos Preview, to a small number of trusted organisations. Washington had previously forced Anthropic to suspend access to its publicly released models in June and has since established a voluntary process for vetting powerful AI systems before they are launched.

The announcement followed a meeting at the White House where Donald Trump hosted top technology executives, including Google’s Sundar Pichai and Anthropic’s Dario Amodei. The group signed a voluntary accord pledging to police the risks associated with their own AI systems. Cybersecurity experts have expressed concern that state-of-the-art technology could be exploited to hack banks, hospitals and government systems. Google claims Argon excels at complex tasks in software engineering, legal and financial work, and cyber-defense, with a leading ability to find and fix critical software flaws.

Early testers used the model to uncover a flaw in software used by hospitals around the world that exposed sensitive personal information, a vulnerability that other advanced models had missed. Google said Argon is designed to refuse requests that could help carry out cyber-attacks or develop chemical, biological or nuclear weapons. The company is also monitoring the model’s reasoning to stop it from straying beyond user intent, a risk researchers call misalignment. This issue has gained urgency after OpenAI disclosed in July that two of its models broke out of a sealed test environment during a cybersecurity evaluation and hacked into the servers of the AI company Hugging Face.