Google launches Gemini 4 Carbon, the AI that finds and fixes security vulnerabilities

Google has officially unveiled Gemini 4 Carbon, its new and most powerful artificial intelligence model. Announced on October 28th, the system has taken the top spot on the Intelligence Index, an independent benchmark that evaluates the capabilities of AI models, surpassing major competitors like Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra.

CEO Sundar Pichai shared the result on X, while Google described Carbon as its most advanced model to date. The main innovation lies in cybersecurity: Carbon is capable of identifying software vulnerabilities and autonomously writing code to fix them. Currently, the model is only available to a select group of security experts, with a broader release for businesses, developers, and consumers planned for the future without specific dates.

Beyond security, Google has enhanced Carbon's capabilities in executing complex and extended tasks ("long-term agents") and in response consistency, reducing the risk of the model taking unforeseen alternative paths to complete an instruction.

The race to develop increasingly powerful and specialized artificial intelligences has been underway for years, with Google, OpenAI, and Anthropic as the main contenders. This competition originated from the revolution in large-scale language models, which, starting from the late 2010s, demonstrated surprising capabilities in understanding and generating human language. The initial goal was to create increasingly useful digital assistants, but the challenge quickly shifted towards models capable of performing complex tasks autonomously.

The cybersecurity sector has become crucial because AIs themselves, if poorly controlled or manipulated, can pose a risk. The problem of "prompt injections" (hidden instructions that induce the AI into unwanted behaviors) and the need for systems that rigorously respect received directives have pushed Google to develop a specific model. Gemini 4 Carbon derives from the previous Gemini 4 Argon and was trained with a particular focus on cybersecurity, testing its resistance to manipulation attempts.

This evolution responds to the growing demand for reliable AI in critical environments, where errors or unexpected behaviors could have serious consequences. The gradual release, first to experts and then to the public, follows an established industry practice to test the safety and effectiveness of new models in controlled contexts before mass adoption.