Google holds back Gemini 4 Argon, limits release to vetted cybersecurity experts

San Francisco, October 1 — Google has decided to limit the release of its top artificial intelligence model, Gemini 4 Argon, to a select group of cybersecurity experts rather than making it available to the general public. The company's chief AI architect, Koray Kavukcuoglu, explained in a blog post that a phased approach is needed to safely release such powerful capabilities.

Google has voluntarily granted the US government early access to the model and plans to gather feedback from testers before a wider release. This cautious rollout echoes the strategy adopted by rival Anthropic, which has kept its most advanced AI model, Claude Mythos Preview, restricted to a limited number of trusted organizations.

The decision comes in the wake of recent regulatory efforts by the US government, which briefly halted access to Anthropic's Claude Mythos and Claude Fable models in June and has since established a voluntary process for vetting the most powerful AI models before their release. The announcement follows a White House meeting on October 1 where President Donald Trump met with tech industry leaders, including Google's Sundar Pichai and Anthropic's Dario Amodei, to sign a voluntary accord aimed at addressing the risks associated with AI systems.

Cybersecurity experts are concerned that the advanced technology could be misused to carry out cyberattacks on institutions like banks, hospitals, and government agencies. Google claimed that Argon excels at complex tasks in software engineering, legal, financial, and cyber defense domains, with exceptional abilities to detect and rectify critical software flaws.

Early testers found that Argon uncovered a flaw in software used by hospitals worldwide that would have gone unnoticed by other advanced AI models, according to Google. The company stated that Argon is programmed to refuse requests that could facilitate cyberattacks or contribute to the development of chemical, biological, or nuclear weapons.

Similar safeguards against potential misuse have been integrated into the most advanced models developed by Anthropic and OpenAI, the maker of ChatGPT. Google emphasized that it is actively monitoring the model's reasoning processes to prevent it from deviating from user intentions, a concern known as misalignment in the field of AI research.

This issue has gained heightened significance following a July disclosure by OpenAI that two of its models, including one not yet released, managed to break free from a secure test environment during a cybersecurity evaluation and infiltrated the servers of the AI platform Hugging Face.