Google launches Gemini 4 Carbon with a ‘significant leap’ in cybersecurity

Your coding assistant is now materially better at finding and fixing security flaws, and you can do nothing to change it.

Google has launched its latest advanced AI model, Gemini 4 Carbon, featuring significant improvements in cybersecurity capabilities. Announced on Wednesday, the model reportedly leads on the Artificial Analysis Intelligence Index (v5.0) benchmark, surpassing competitors like Claude Opus 5.5 and GPT-6 Astra. Gemini 4 Carbon builds upon its predecessor, Argon, released earlier this month, with enhanced abilities in finding vulnerabilities and generating secure code. It also excels in discovering real-world bugs and creating exploits, according to performance on the SEC-bench Pro and ExploitGym benchmarks. Initially available to select cybersecurity defense companies through Google's Fairwind Program, the model is designed for extended, complex tasks, demonstrating increased resilience against prompt injection and offering greater transparency by reducing unsupported claims. The pricing remains competitive, matching Argon's introductory rates of $2 per million input tokens and $10 per million output tokens.

The release of Gemini 4 Carbon signals a heightened focus on AI's role in cybersecurity, with Google positioning its model as a leader in vulnerability detection and secure code generation. This advancement suggests a market trend toward more robust AI-driven security solutions, potentially setting new industry standards for how AI models are evaluated and utilized for defense. The model's superior performance on specialized benchmarks like SEC-bench Pro and ExploitGym indicates a maturing capability in practical security applications, moving beyond theoretical benchmarks.

The availability of Gemini 4 Carbon, even initially to a select group, points to the increasing integration of advanced AI into critical infrastructure and defense sectors. Companies that adopt this technology could see a significant advantage in their security posture, while those not leveraging such tools may face increasing cybersecurity risks. The emphasis on longer task completion and resistance to attacks like prompt injection suggests future AI models will be more autonomous and reliable for complex, multi-step security operations.