Google is holding back its most powerful artificial intelligence system from the wider public, limiting access to Gemini 4 Argon to a vetted group of cybersecurity experts amid concerns that hackers could exploit its capabilities.
“Safely releasing frontier capabilities at this level requires a phased approach,” Koray Kavukcuoglu, Google’s chief AI architect, wrote in a blog post announcing the model.
- Advertisement -
The company said it is also granting the US government early access and will use feedback from the initial testers to guide any broader release.
That measured strategy echoes rival Anthropic, which has similarly restricted its most advanced system, Claude Mythos Preview, to a select group of trusted organizations.
Google’s announcement came one day after President Donald Trump welcomed leading technology executives — including Google chief Sundar Pichai and Anthropic’s Dario Amodei — to the White House, where they signed a voluntary agreement promising to oversee 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
Cybersecurity specialists have warned that technology at the cutting edge could be turned against banks, hospitals and government networks.
Google said Argon can handle demanding work in software engineering, law, finance and cyber defense. The company said the model is particularly capable of identifying and repairing critical weaknesses in software.
During early testing, Argon detected a flaw in software used by hospitals worldwide that had left sensitive personal information exposed, Google said — a vulnerability that other advanced models failed to find.
Google said Argon has been trained to reject requests that could facilitate cyberattacks or the creation of chemical, biological or nuclear weapons.
Anthropic and OpenAI, the company behind ChatGPT, have introduced comparable safeguards in their own most advanced systems.
Google also said it is tracking Argon’s reasoning to prevent the model from moving beyond a user’s intended goal, a danger researchers describe as misalignment.
That concern has become more pressing since OpenAI disclosed in July that two of its models, including one that had not yet been released, escaped a sealed testing environment during a cybersecurity assessment and hacked into the servers of AI company Hugging Face.
