Google Launches Gemini 4 Carbon: Advanced AI for Vulnerability Detection and Secure Code Generation

Introducing Gemini 4 Carbon AI Model

According to НВ — Техно: On October 29 at 6:32 PM, Google unveiled its latest artificial intelligence model, Gemini 4 Carbon. Designed to identify software vulnerabilities, generate fixes, and tackle complex engineering tasks over extended periods, this AI demonstrates cutting-edge capabilities. Gemini 4 Carbon outperformed competitors in three key benchmarks: SEC-bench Pro, CWE-bench, and ExploitGym, while also featuring enhanced security safeguards.

Benchmark Performance Highlights

In the SEC-bench Pro test, which measures an AI's skill in detecting vulnerabilities in real-world software, Gemini 4 Carbon scored 88.2%. The CWE-bench evaluation, focused on the model's ability to repair security flaws, returned a 71% success rate. For the ExploitGym test, assessing controlled exploitation attempts to verify vulnerabilities, the AI achieved 54.8%. Across all these evaluations, Gemini 4 Carbon led its peers.

Internal assessments reveal that Gemini 4 Carbon more than doubled the success rate of producing secure code compared to its predecessor, Gemini 4 Argon. According to Chrome's engineering team, Carbon surpasses another top-tier model by over tenfold in the quality of vulnerability patches it generates. Engineered for prolonged autonomous operation, Gemini 4 Carbon can handle multi-step tasks, utilize diverse tools, investigate root causes of software bugs, and develop fixes for large codebases. Google teams observed the model working on intricate engineering challenges continuously over several days.

Gemini 4 Carbon secured first place on the Artificial Analysis Intelligence Index version 5.0, earning a score of 59 points. In the WebDev Arena, its rating increased by 89 Elo points to 1768, ranking it fourth overall. The AI also showed improvements in scientific problem-solving, data handling, and web application development. Internal tests demonstrated roughly a 50% reduction in unsupported claims compared to Argon, and serious consistency warnings dropped by nearly 80% in programming environments used by Google teams. In the Gray Swan test, which assesses resistance to indirect prompt-based attacks, Gemini 4 Carbon recorded zero successful attacks, outperforming Argon's 0.7% vulnerability.

Initially, access to Gemini 4 Carbon will be granted to a select group of cybersecurity experts, with plans to expand availability to trusted partners through Google's Fairwind program. Google is also participating in a voluntary pre-deployment review process with the U.S. government. Wider access for developers, businesses, and end users is planned but without a set release date. During the early rollout, usage pricing is set at $2 per million input tokens and $10 per million output tokens, with a 95% discount applied to cached input tokens. The limited launch reflects concerns over dual-use risks, as the technology could be leveraged both to enhance security and potentially to facilitate attacks.

Google's Gemini 4 Carbon marks a major advancement in AI-driven cybersecurity. Its capabilities in vulnerability detection and automated patch creation position it as a crucial tool for strengthening software defenses.

As Google continues to innovate in the AI landscape, the recent introduction of Gemini AI Agent highlights its commitment to enhancing task automation and corporate system integration. This development complements the capabilities of Gemini 4 Carbon, showcasing how Google is streamlining processes across various domains to improve efficiency and productivity.