Google has restricted public access to its most powerful artificial intelligence model, Gemini 4 Argon, allowing only a vetted group of cybersecurity experts to test it over concerns that hackers could misuse the technology.
“Safely releasing frontier capabilities at this level requires a phased approach,” Google’s chief AI architect, Koray Kavukcuoglu, said in a blog post announcing the model.
Google said it was voluntarily giving the US government early access to Argon and would gather feedback from testers before considering a wider release.
The cautious rollout mirrors the approach of rival Anthropic, which has also restricted access to its most advanced model, Claude Mythos Preview, to a limited number of trusted organisations.
Washington briefly forced Anthropic to suspend access to its publicly released Claude Mythos and Claude Fable models in June. The US has since established a voluntary process for vetting the most powerful AI models before their release.
The announcement came a day after President Donald Trump hosted top technology executives, including Google chief executive Sundar Pichai and Anthropic’s Dario Amodei, at the White House, where they signed a voluntary agreement to address risks associated with their AI systems.
Cybersecurity experts have raised concerns that advanced AI could be used to attack banks, hospitals and government systems.
Google said Argon excels at complex tasks in software engineering, legal and financial work, as well as cyber defence. The company said the model has a leading ability to identify and fix critical software vulnerabilities.
Early testers reportedly used Argon to uncover a flaw in software used by hospitals around the world that exposed sensitive personal information, which Google said other advanced models had failed to detect.
Google said Argon was designed to refuse requests that could assist cyberattacks or the development of chemical, biological or nuclear weapons.
Anthropic and OpenAI have implemented similar safeguards in their most advanced models.
Google also said it was monitoring Argon’s reasoning to prevent it from acting beyond users’ intentions, a risk researchers refer to as misalignment.
The concern has gained renewed attention since OpenAI disclosed in July that two of its models, including one that had not yet been released, escaped a sealed test environment during a cybersecurity evaluation and accessed the servers of AI company Hugging Face.