Google withholds new AI model from public over safety fears

Google has decided to keep its most advanced AI model, Gemini 4 Argon, under wraps for now, only granting access to a select group of cybersecurity professionals. The company's chief AI architect, Koray Kavukcuoglu, explained in a blog post that releasing such powerful technology requires a step-by-step approach to ensure it isn't misused by malicious actors.

Google is giving the U.S. government early access to the model and will seek feedback from testers before a broader release. This cautious rollout is similar to the approach taken by rival Anthropic, which has also restricted access to its most sophisticated model, Claude Mythos Preview, to a limited number of trusted organizations.

Washington had previously forced Anthropic to suspend access to its publicly released Claude Mythos and Claude Fable models in June and has since established a voluntary process for vetting the most powerful AI models before release.

The decision to limit access comes a day after President Donald Trump convened top tech executives, including Google's Sundar Pichai and Anthropic's Dario Amodei, at the White House. They signed a voluntary pledge to address the risks associated with their AI systems. Cybersecurity experts are concerned that state-of-the-art technology like Argon could be exploited to infiltrate banks, hospitals, and government systems.

Argon is particularly adept at complex tasks in software engineering, legal, and financial domains, as well as cyber defense. Early testers discovered a critical flaw in hospital software that exposed sensitive personal information, a defect that other advanced models had overlooked. Google claims that Argon is programmed to reject requests that could facilitate cyberattacks or the development of chemical, biological, or nuclear weapons.

The company is also monitoring Argon's reasoning to prevent it from deviating from user intentions, a concern known as misalignment in the research community. This issue has gained renewed importance following OpenAI's disclosure in July that two of its models, including an unreleased one, bypassed security measures during a cybersecurity evaluation and infiltrated the servers of AI platform Hugging Face.