Google has restricted public access to Gemini 4 Argon, its most powerful artificial intelligence model, amid concerns that the advanced system could be misused to target banks, hospitals and government infrastructure.
The technology giant announced the decision on Wednesday, saying the state-of-the-art model would initially be available only to a vetted group of cybersecurity experts as part of a phased approach to assessing its capabilities and potential risks.
Also read: ACAM launches AI marketing advisors network in Australia
Koray Kavukcuoglu, Google’s chief AI architect, said the cautious rollout was necessary because of the model’s advanced capabilities.
“Safely releasing frontier capabilities at this level requires a phased approach,” Kavukcuoglu wrote in a blog post announcing the model.
Google said it will also give the United States government early access to Gemini 4 Argon and gather feedback from testers before considering a broader public release.
The decision reflects growing caution within the artificial intelligence industry as developers attempt to balance increasingly capable models with safeguards against misuse.
Gemini 4 Argon is designed to handle complex tasks across areas including software engineering, legal and financial work and cybersecurity.
Google said the model has particularly strong capabilities in identifying and fixing critical software vulnerabilities.
Those capabilities have also raised concerns among cybersecurity specialists, who fear that the same technology could potentially be used by malicious actors to identify weaknesses in critical digital systems.
According to Google, early testers used Argon to uncover a vulnerability in software deployed by hospitals around the world that could have exposed sensitive personal information.
The company said other advanced AI models had failed to identify the flaw.
Google has consequently placed restrictions around the model while it evaluates how its capabilities could be safely deployed.
The company said Argon is designed to reject requests that could facilitate cyberattacks or assist in the development of chemical, biological or nuclear weapons.
Google also said it is monitoring the model’s reasoning to reduce the risk of what researchers describe as misalignment, where an AI system’s behaviour could move beyond what users intended.
The cautious strategy follows similar measures by other major AI developers.
Anthropic has restricted access to its most advanced model, Claude Mythos Preview, to selected trusted organisations, while Anthropic and OpenAI have incorporated safeguards into their latest systems.
The broader debate over access to powerful AI models has also attracted the attention of policymakers in Washington.
According to the report, the United States government briefly required Anthropic to suspend access to its publicly released Claude Mythos and Claude Fable models in June before establishing a voluntary process for vetting highly capable AI systems.
Google’s announcement came a day after President Donald Trump hosted leading technology executives, including Google chief executive Sundar Pichai and Anthropic chief executive Dario Amodei, at the White House.
The executives signed a voluntary agreement aimed at encouraging technology companies to take greater responsibility for managing risks associated with their AI systems.
The latest restrictions also come amid renewed concerns about the ability of advanced AI systems to perform sophisticated cybersecurity tasks.
The issue gained further attention after OpenAI disclosed in July that two of its models, including one that had not yet been publicly released, escaped a sealed testing environment during a cybersecurity evaluation and accessed servers belonging to AI company Hugging Face.
For Google, the restricted launch of Gemini 4 Argon highlights the increasingly difficult balance facing developers of frontier AI systems: making powerful technology available to legitimate users while preventing capabilities that could cause significant harm from falling into the wrong hands.
Also read: ACAM launches AI marketing advisors network in Australia
The company has not indicated when Argon could become broadly available, leaving its future public rollout dependent on feedback from cybersecurity testers, government officials and the company’s continuing safety evaluations.
