Google introduces Gemini 4 Carbon - an AI model for detecting and fixing software vulnerabilities

Google has unveiled Gemini 4 Carbon, a new AI model focused on cybersecurity, software development, and complex engineering tasks.

The model scored 88.2% on SEC-bench Pro (vulnerability detection), 71% on CWE-bench (vulnerability remediation), and 54.8% on ExploitGym (testing the exploitability of vulnerabilities).

According to Google, Carbon more than doubled Argon's performance in secure code generation. In Chrome team's testing, it also delivered 10 times better results than a competing model in generating high-quality security updates.

Unlike conventional AI assistants, Gemini 4 Carbon is designed to handle long-running, multi-step tasks, analyze large codebases, identify bugs, and develop fixes. It also demonstrates improved capabilities in reasoning, scientific tasks, and web data processing.

Google says it has strengthened the model's safety mechanisms: unsupported claims have been reduced by approximately 50% compared with Argon, while critical misalignment alerts have dropped by nearly 80%.

Access to Gemini 4 Carbon will initially be limited to a select group of cybersecurity experts.