Google unveils Gemini 4 Carbon, an AI that can find software vulnerabilities

Google DeepMind has launched Gemini 4 Carbon, a new artificial intelligence model that can identify software vulnerabilities, test ways to exploit them and generate security fixes. Announced on Wednesday, 28 October, the model is described by Google as its most advanced AI for programming and technical problem-solving.

Unlike typical chatbots, Carbon can run multi-day tasks, including diagnosing errors and modifying code to fix them. Google reports strong results on security benchmarks: 88.2% on SEC-bench Pro, which tests vulnerability discovery in real applications, 54.8% on ExploitGym for exploitation and 71% on CWE-bench for generating fixes. These scores were achieved without the usual safety mechanisms, and do not mean the model can compromise any system.

Because the technology could be misused, Google is rolling it out gradually. Initial access is limited to a selected group of cybersecurity specialists. It will expand to trusted partners through the Fairwind programme as Google evaluates protections against misuse. The company also notes that in a benchmark for hidden-instruction attacks, the model had a 0% success rate, though that is not a guarantee.

Google plans to extend availability to developers, companies and general users but has not set a date. Introductory pricing is set at $2 per million input tokens and $10 per million output tokens. The model is said to generate secure code at more than twice the success rate of Gemini 4 Argon, released two weeks earlier.

For Romanian developers and businesses focused on software security, Gemini 4 Carbon could become a powerful tool once access widens.