Google has introduced its new advanced artificial intelligence model, Gemini 4 Argon, but has decided against making it publicly available for now, citing concerns over potential misuse.
The company said the model will initially be provided only to a carefully selected group of cybersecurity experts and partners. Google’s chief AI architect Koray Kavukcuoglu said a phased rollout was necessary to safely introduce AI capabilities at this level.
According to Google, Gemini 4 Argon is larger and more capable than its earlier advanced “Pro” models. The company described it as its most powerful model so far for complex workloads, particularly in software development and cybersecurity.
Google said Argon delivered competitive results against leading AI models from OpenAI and Anthropic across several coding and cybersecurity benchmarks. However, the company acknowledged that its model did not outperform rivals on every test, including some coding-related evaluations.
The technology is designed to handle complex software engineering, legal, financial and cybersecurity tasks. Google said one of its notable capabilities is identifying and fixing serious software vulnerabilities.
During early testing, Argon reportedly identified a security flaw in software used by hospitals that could have exposed sensitive personal information. Google said other advanced AI models involved in testing had failed to detect the vulnerability.
Google said Gemini 4 Argon has safeguards designed to reject requests that could facilitate cyberattacks or assist in the development of chemical, biological or nuclear weapons.
The company said it is working with selected cybersecurity organisations and participating in a US government initiative focused on evaluating powerful AI models before wider release.
Google’s cautious approach comes amid growing concerns among cybersecurity researchers that increasingly capable AI systems could be exploited to target banks, hospitals and government networks.
The company is also monitoring Argon’s reasoning processes to reduce the risk of the model acting beyond a user’s intended instructions, an issue commonly referred to as AI misalignment.
The announcement came shortly after US President Donald Trump met senior technology executives, including Google CEO Sundar Pichai and Anthropic CEO Dario Amodei, at the White House. The executives agreed to a voluntary framework aimed at managing risks associated with advanced AI systems.
The decision to restrict Argon’s initial access follows a similar strategy adopted by rival AI companies, which have also limited access to their most advanced systems to selected organisations.
Concerns over AI cybersecurity have intensified following reports of advanced models demonstrating the ability to exploit vulnerabilities during controlled security tests.
Google has also decided not to proceed with the previously planned release of Gemini 3.5 Pro, which had initially been expected to launch earlier.
Meanwhile, Google has undergone significant changes within its AI division, including leadership changes at DeepMind and departures involving members of its Gemini team.
The delays surrounding Google’s latest AI models have added pressure on the company as competitors continue releasing increasingly advanced systems.
Despite the competitive pressure, Pichai has rejected suggestions that Google is falling behind in the AI race. In recent months, Google has increasingly highlighted Gemini’s cost efficiency alongside its capabilities as it competes with rival AI developers.
