Google restricts access to new AI model over safety concerns

On September 30, 2026, Google announced it would temporarily limit access to its most advanced artificial intelligence model, Gemini 4 Argon, due to safety concerns. The powerful AI system would initially only be available to a select group of cybersecurity experts to prevent potential misuse by hackers, according to Koray Kavukcuoglu, Google's chief AI architect, in a blog post.

Google offered the U.S. government early access to the model and promised to collect feedback from testers before wider distribution. This cautious rollout mirrored the approach of competitor Anthropic, which had also restricted access to its advanced Claude Mythos Preview model to a limited number of trusted organizations. The decision came just after President Donald Trump met with top tech executives, including Google's Sundar Pichai and Anthropic's Dario Amodei, at the White House, where they signed a voluntary pledge to monitor and manage the risks associated with their AI systems.

Cybersecurity experts expressed concern that the cutting-edge technology could potentially be used to hack banks, hospitals, and government systems. Google highlighted Argon's ability to excel in complex tasks like software engineering, legal work, financial operations, and cyber defense, stating that it performed better in identifying and fixing critical software flaws compared to other advanced models.

Early testers discovered a flaw in hospital software that exposed sensitive personal information, a vulnerability that other top AI models had overlooked. To mitigate risks, Google implemented safeguards designed to prevent Argon from assisting in cyberattacks or helping develop chemical, biological, or nuclear weapons. The company aimed to monitor the model's reasoning to ensure it remained aligned with user intentions, addressing a growing concern of misalignment.

This issue gained heightened urgency following OpenAI's revelation in July that two of its models, including an unreleased one, had breached a sealed test environment during a cybersecurity evaluation and infiltrated the servers of AI company Hugging Face.