Google Unveils Gemini 4 Carbon With Advanced AI Cybersecurity Capabilities

Google has unveiled Gemini 4 Carbon, its latest frontier artificial intelligence model, marking another major development in the company’s efforts to compete with leading AI developers. Announced on October 28, 2026, the model builds on Gemini 4 Argon, which Google introduced earlier in the month. According to Google DeepMind, Carbon delivers significant improvements in complex coding tasks, long-running autonomous workflows and cybersecurity operations. The company describes it as its strongest model so far, particularly for identifying software vulnerabilities, validating security weaknesses and developing solutions to fix them. The announcement comes as technology companies increasingly develop AI systems capable of performing complicated technical tasks with less direct human supervision.

One of the model’s most notable features is its ability to handle long-running engineering projects. Google says Carbon can work through complex problems across large software codebases, investigate difficult bugs and develop fixes that may require several stages of reasoning. The company reports that its internal teams have successfully used the model for engineering tasks that continued for days, including identifying the underlying causes of subtle software failures. Carbon also recorded an improvement of 89 Elo points on WebDev Arena compared with Argon, reaching fourth position with a score of 1,768. These capabilities could make the model valuable to software developers, engineering teams and organisations seeking to automate parts of their development processes.

Cybersecurity is the central focus of Gemini 4 Carbon. Google reports that the model achieved 88.2% on SEC-bench Pro, a benchmark designed to evaluate vulnerability discovery in real-world software. It also recorded 54.8% on ExploitGym, which assesses the ability to validate whether security weaknesses can be exploited, and 71% on CWE-bench for identifying and fixing software vulnerabilities. Google further claims that Carbon delivers more than twice the success rate of Argon when generating secure code in its internal evaluations. These results suggest that the model could help cybersecurity professionals identify weaknesses earlier, assess potential threats and develop more effective software patches. However, the reported figures are based on specific testing conditions and do not guarantee identical performance in every real-world security environment.

Google has also emphasised improvements in safety and reliability. The company says Carbon was trained to reduce fabricated claims, avoid taking actions beyond user instructions and resist malicious instructions hidden inside external information. According to its internal evaluations, the model reduced high-severity misalignment alerts by almost 80% in live coding environments. It also recorded a 0% attack success rate on the evaluated Gray Swan Indirect Prompt Injection benchmark, compared with Argon’s previously reported 0.7%. Google says these improvements are supported by multiple security measures, including model training, real-time misuse protections, prompt-injection resistance and monitoring. The company acknowledges that powerful cybersecurity models require careful deployment because the same technical capabilities that help defenders identify vulnerabilities could potentially be misused by malicious actors.

Unlike a conventional public launch, Google is initially restricting access to Gemini 4 Carbon. The company says it will first provide the model to selected cybersecurity professionals before gradually expanding access through its Fairwind programme to trusted partners. Google is also participating in the United States government’s voluntary process for evaluating advanced AI models before release. The introductory API price is $2 per million input tokens and $10 per million output tokens, matching the initial pricing of Gemini 4 Argon. Google says these prices will eventually increase to $4 per million input tokens and $20 per million output tokens after the introductory period. Wider availability for developers, businesses and ordinary consumers is expected later.

The launch strengthens Google’s position in the increasingly competitive market for advanced AI systems. OpenAI and Anthropic are also investing heavily in models capable of autonomous coding, software development and cybersecurity analysis. By introducing Gemini 4 Carbon shortly after Argon, Google is demonstrating its intention to accelerate development while targeting specialised areas where advanced AI could have significant commercial and security applications. Nevertheless, the model’s real-world impact will depend on independent evaluations, practical performance and the effectiveness of its safety controls. If its reported capabilities translate into reliable results outside Google’s testing environments, Gemini 4 Carbon could become an important tool for software engineering and cybersecurity teams worldwide.