October 7, 2026, 8:07 am | Read time: 2 minutes
Google has developed an AI model called Gemini 4 Argon, designed to detect even complex security vulnerabilities. However, instead of a public launch, only experts and authorities are initially allowed to test the technology.
Access limited to selected experts
With Gemini 4 Argon, Google is pursuing an unusual strategy. While new AI models are typically quickly released to developers or the public, Argon remains behind closed doors for now. Access is initially granted only to selected specialists in the field of cybersecurity.
According to Google, the background involves potential risks associated with the system’s exceptional capabilities. Therefore, the technology is to be tested under controlled conditions first. Authorities in the U.S. are also allowed to evaluate the model. Only when testers no longer see significant concerns might Google make the AI more widely available.
ChatGPT Goes Rogue and Conducts Cyberattack Independently
Gemini Infiltrated Real Corporate Networks
AI to detect critical security vulnerabilities
According to the company, Gemini 4 Argon is among the most advanced AI systems Google has developed to date. The model is said to be particularly impressive in areas such as software development, cybersecurity, law, and finance. One of its key strengths is finding hard-to-detect vulnerabilities in programs.
Google reported that in initial tests, the AI identified a security flaw in a globally used hospital software. The vulnerability could have endangered sensitive patient data. Other modern AI systems had apparently not detected the error before. For security researchers, such a capability could be a valuable tool in the future.
Also of interest: Google Gemini infiltrated foreign company networks
Between protective tool and potential threat
However, this very capability is also causing discussions. Experts fear that highly developed AI systems could be used not only for defense but also for attacks on digital infrastructures. Institutions such as hospitals, banks, or authorities are considered potential targets.
Google emphasizes that Gemini 4 Argon has been explicitly trained to reject requests that could assist in hacking computer systems or provide information on building biological, chemical, or nuclear weapons. Whether these security mechanisms function adequately is to be determined in the current testing phase. Until then, Google’s most powerful AI remains reserved for a small circle of selected users.