Google has opted for a limited rollout of its most advanced artificial intelligence model, Gemini 4 Argon, restricting access to a selected group of cybersecurity specialists over concerns that hackers could misuse its capabilities.
The company said the phased release would allow it to assess the model’s risks and gather feedback before making it available to the wider public.
Google’s chief AI architect, Koray Kavukcuoglu, said the approach was necessary because of the model’s advanced capabilities.
The company is also voluntarily providing the United States government with early access to Argon as part of efforts to evaluate the technology before a broader release.
The move follows a similar strategy by AI rival Anthropic, which has limited access to its most advanced model, Claude Mythos Preview, to a small group of trusted organisations.
Concerns have grown among cybersecurity experts that highly capable AI systems could potentially be used to target banks, hospitals, government networks and other critical infrastructure.
Google said Argon has demonstrated strong performance in software engineering, legal and financial tasks, as well as cybersecurity. The company said the model can identify and repair serious software vulnerabilities.
During early testing, Argon reportedly helped researchers discover a security flaw in software used by hospitals globally, which could have exposed sensitive personal information. Google said other advanced AI models had failed to identify the vulnerability.
To reduce potential misuse, Google said Argon has been built to reject requests that could facilitate cyberattacks or the development of chemical, biological or nuclear weapons.
Similar safeguards have also been incorporated into advanced AI systems developed by Anthropic and OpenAI.
Google said it is additionally monitoring Argon’s reasoning processes to ensure the model remains within the scope of what users intend, addressing concerns researchers describe as AI misalignment.
The issue has gained further attention after OpenAI disclosed in July that two of its models, including one that had not yet been publicly released, escaped a controlled testing environment during a cybersecurity assessment and accessed the servers of AI company Hugging Face.
