Google’s most advanced model is here, but only “trusted cyber defenders” and the US government get to use it
Gemini 4 Argon can find critical vulnerabilities all on its own.
Image by Cybernews.
Follow us
- Google says Gemini 4 Argon outperforms rival models across coding, knowledge work, and cybersecurity benchmarks.
- The model can find, verify, and patch critical software flaws without human input, according to Google.
- Google already uses Argon for safer code migration and complex quantum computing research tasks.
- Under Trump’s order, US agencies get access to powerful AI models 30 days before other trusted partners.
Key Takeaways by nexos.ai, reviewed by Cybernews staff.
Google has announced its most advanced model yet, Gemini 4 Argon. The AI model is already being used internally and by “trusted partners.” But before it can go public, it has to go through the Trump Administration.
Google is rolling out Gemini 4 Argon, its frontier performance model, which claims to outperform competing models such as OpenAI’s Astra and Anthropic’s Fabel 5.1 and Opus 5.5 across a range of benchmarks.
Google built its Gemini 4 Argon model with “frontier-level capabilities on coding, knowledge work, and cybersecurity defense,” positioning the model as one of its most powerful to date.
The model is already being used at Google to migrate C/C++ codebases to the more secure language Rust, and to assist the tech giant’s researchers with complex tasks in quantum computing.
Argon is currently outperforming models in the Vals Index vibecoding benchmark, demonstrating its ability to create applications from scratch using human language.
The Vals Index also measures economic impact across finance, coding, legal, and tax work, weighted against its contribution to the US Gross Domestic Product, and Argon is in the lead, according to Google.
The model is apparently extremely good at finding, validating, and patching critical software vulnerabilities without human input.
AI models capable of patching vulnerabilities have also been known to exploit them, as demonstrated by OpenAI and Google itself following the Hugging Face hack and Gemini hack.
But the Trump Administration is attempting to curb these incidents via an executive order that the president signed in June this year.
Stay updated with our latest stories and follow us on social media
Be the first to discover new stories, ideas, and updates from our team.
Powerful AI models must go through Trump first
Google is voluntarily working with the US government to ensure that powerful AI models are rolled out responsibly and the technology doesn’t go rogue as it has in the past.
AI companies that agree to this must provide federal government organizations with unbridled access to what the White House calls “covered frontier models” 30 days before planned release to other trusted partners.
It’s unclear who these “trusted partners” are, as Google set up its own program separate from the US government called Fairwind.
This encompasses roughly 650 of Google’s global partners, like CrowdStrike and Accenture, who could receive early access to Argon.
In September, President Donald Trump asked OpenAI and Anthropic not to share their new AI models with the UK's flagship testing agency before the US government performed a risk assessment, according to Bloomberg.
