Google Limits Access to New AI Model Amid Rising Safety Concerns

Google is keeping its most powerful artificial intelligence system out of public hands for now, opting instead for a tightly controlled debut designed to deny hackers an easy new tool.

The company said Gemini 4 Argon will be provided only to a vetted group of cybersecurity experts as Google weighs how to prevent misuse.

“Safely releasing frontier capabilities at this level requires a phased approach,” wrote Koray Kavukcuoglu, Google’s chief AI architect, in a blog post announcing the model.

Google said it is also voluntarily granting the US government early access and will use feedback from initial testers before deciding on a broader release.

The guarded rollout echoes rival Anthropic’s strategy: it has kept its most advanced model, Claude Mythos Preview, limited to a small circle of trusted organizations rather than opening it to the wider public.

Washington briefly forced Anthropic to halt access to its publicly released Claude Mythos and Claude Fable models in June, and the government has since established a voluntary vetting process for the most powerful AI systems ahead of release.

Google’s announcement landed a day after President Donald Trump convened leading tech executives at the White House — including Google chief Sundar Pichai and Anthropic’s Dario Amodei — where they signed a voluntary accord committing to police the risks posed by their own AI systems.

Google CEO Sundar Pichai was among a number of tech executives hosted by US President Donald Trump yesterday

Security specialists have warned that cutting-edge models, if widely available, could be turned into weapons for cybercrime, enabling attacks on banks, hospitals and government systems.

Google said Argon is particularly strong at complex work across software engineering, legal and financial tasks, and cyber defense, highlighting what it called a leading ability to identify and repair critical software flaws.

In one early test, Google said Argon found a vulnerability in software used by hospitals around the world that exposed sensitive personal information — a weakness other advanced models had failed to detect.

Google said it built Argon to refuse requests that might assist cyberattacks or contribute to the development of chemical, biological or nuclear weapons.

Anthropic and ChatGPT-maker OpenAI have put comparable safeguards into their most advanced models.

Google said it is also tracking the model’s reasoning to prevent it from drifting beyond what users intend — a failure mode researchers describe as misalignment.

Concerns about that risk have intensified since OpenAI reported in July that two of its models, including one not yet released, escaped a sealed test environment during a cybersecurity assessment and hacked into servers belonging to AI company Hugging Face.