Google Holds Back New AI Model, Limits Gemini 4 Argon To Security Experts

Markus Habila
Editor

Google Holds Back New AI Model, Limits Gemini 4 Argon To Security Experts
Google has withheld its latest artificial intelligence model, Gemini 4 Argon, from general public access, limiting its initial use to vetted cybersecurity experts over concerns about potential misuse.
The company announced the decision on Wednesday, saying the model’s advanced capabilities required a phased release to allow researchers and security specialists to assess its potential risks.
Google Chief AI Architect, Koray Kavukcuoglu, said the company would gather feedback from early users before deciding whether to make Argon more widely available.
The company has also granted the United States government early access to the model as part of the testing process.
Google said Gemini 4 Argon was developed to handle complex tasks in areas including software engineering, legal and financial work and cybersecurity. It added that the model has a particularly strong capability to identify and fix critical software vulnerabilities.
According to the company, early testers used Argon to discover a vulnerability in software deployed by hospitals globally, which could have exposed sensitive personal information. Google said the flaw had not been detected by other advanced AI models.
The restricted release reflects growing concerns over the possibility that highly capable AI systems could be used by malicious actors to conduct sophisticated cyberattacks against banks, hospitals, government agencies and other critical infrastructure.
Google said Argon contains safeguards designed to reject requests that could assist cyberattacks or the development of chemical, biological or nuclear weapons.
It is also monitoring the model’s reasoning to reduce the risk of what researchers describe as “misalignment”, where an AI system behaves in ways that go beyond the user’s intended instructions.
The move follows similar restrictions by other AI developers. Anthropic has limited access to its advanced Claude Mythos Preview model to a small number of trusted organisations.
The issue of AI security has gained further attention following OpenAI’s disclosure in July that two of its models, including one that had not yet been released publicly, escaped a controlled testing environment during a cybersecurity evaluation and accessed servers operated by AI company Hugging Face.
Google said the controlled rollout of Gemini 4 Argon would enable it to assess the model’s capabilities and security risks before determining whether to expand access to the wider public.