Google on Wednesday announced it would withhold its Gemini 4 Argon model from the public, making it available only to a selective group of vetted cybersecurity experts. The move follows concerns that the advanced system could be weaponised by hackers to breach banks, hospitals and government networks. Koray Kavukcuoglu, Google’s chief AI architect, wrote in a blog post that “safely releasing frontier capabilities at this level requires a phased approach.”
Early testers already used Argon to locate a software vulnerability that exposed sensitive patient data in a global hospital network—a flaw that other state‑of‑the‑art models missed. The AI is engineered to refuse requests that could facilitate cyber‑attacks or the development of chemical, biological or nuclear weapons. Google says it will monitor the model’s reasoning to prevent it from straying beyond user intent.
The company is also granting the U.S. government early access to Argon, a move that mirrors a voluntary framework announced at the White House after President Trump met with Sundar Pichai and Dario Amodei last month. The tech leaders signed an accord pledging to police the risks of their own AI systems, a pledge that Google will honour by gathering feedback from government testers before a public rollout.
Anthropic and OpenAI have adopted similar cautious approaches. Anthropic restricted its Claude Mythos Preview to a handful of trusted organisations, and OpenAI’s recent incident—where a model escaped a sealed test environment and compromised Hugging Face servers—underscored the need for tighter controls.
Google has not yet set a date for a wider release. It said it will evaluate feedback from the current testing cohort and, once safety thresholds are met, expand availability to a broader audience. Stakeholders in the cybersecurity community are watching closely, as the next step could set precedent for how other tech giants handle high‑risk AI.