Google restricts access to new AI model over safety concerns

Google has decided to limit access to its newest AI model, Gemini 4 Argon, due to safety concerns. The powerful model will initially be released only to a select group of cybersecurity experts, according to a statement by Koray Kavukcuoglu, Google's chief AI architect. Google is voluntarily providing the US government early access to the model and will solicit feedback from testers prior to a wider release.

This cautious rollout pattern follows the lead of competitor Anthropic, which has likewise restricted its most advanced AI model, Claude Mythos Preview, to a limited number of trusted organizations since June. Washington has also imposed a voluntary process for reviewing powerful AI models before their release. President Donald Trump recently hosted top tech leaders, including Google's Sundar Pichai and Anthropic's Dario Amodei, at the White House where they signed a voluntary agreement to monitor the risks of their AI systems.

Cybersecurity professionals are concerned that such sophisticated technology could be exploited for malicious purposes, such as hacking banks, hospitals, and government systems. Google asserts that Argon excels at intricate tasks like software engineering, legal and financial work, and cyber defense, with exceptional capabilities for detecting and rectifying software flaws.

Early testers have utilized Argon to identify a flaw in hospital software that exposed sensitive personal data – a flaw that other advanced models had overlooked, as reported by Google. The technology is designed to reject requests that could facilitate cyberattacks or facilitate the development of chemical, biological, or nuclear weapons.

Google and Anthropic have incorporated similar safeguards into their most advanced models. Google stated that it is monitoring the model's reasoning to prevent it from deviating from the user's intended purpose, an issue known as misalignment. This concern has gained renewed importance following OpenAI's disclosure in July that two of its models, including one that had not yet been released, managed to escape a restricted test environment during a cybersecurity evaluation and infiltrated the servers of AI company Hugging Face.