Google launches Gemini 4 Carbon, scoring 88% on cybersecurity vulnerability detection

Google announced Gemini 4 Carbon on October 28, 2026, its latest frontier AI model that ranks first on the Artificial Analysis Intelligence Index with a score of 59. The model achieves 88.2% on cybersecurity vulnerability detection (SEC-bench Pro), 54.8% on exploitation validation (ExploitGym), and 71.0% on patching (CWE-bench), outperforming competitors including Claude Opus 5.5 and GPT-6 Astra. Carbon will roll out first to cybersecurity defenders through the Fairwind Program with introductory pricing at $2 per million input tokens.

Quick Facts

Google announced Gemini 4 Carbon frontier AI model
Model ranked first on Artificial Analysis Intelligence Index v5.0
Model achieved 88.2% on SEC-bench Pro for vulnerability detection
Model achieved 54.8% on ExploitGym for exploitation validation
Model achieved 71.0% on CWE-bench for patching

Google announced Gemini 4 Carbon on October 28, 2026, positioning it as its latest frontier artificial intelligence model with enhanced capabilities in coding, autonomous tasks, and cybersecurity. The model arrives less than two weeks after the release of Gemini 4 Argon, marking an accelerated development cadence that Google attributes to researchers and AI models working together to advance capabilities faster.

On the Artificial Analysis Intelligence Index v5.0, Gemini 4 Carbon ranks first with a score of 59, surpassing Anthropic's Claude Opus 5.5 (58) and OpenAI's GPT-6 Astra (57). The model demonstrates particular strength in cybersecurity benchmarks, achieving 88.2% on SEC-bench Pro for vulnerability detection, 54.8% on ExploitGym for exploitation validation, and 71.0% on CWE-bench for patching. Google states that Carbon is twice as likely to generate secure code compared to its predecessor Argon, and that the Chrome team reports Carbon is 10 times better at generating high-quality security patches than non-Google models. The model also scored 1,769 points on WebDev Arena, ranking fourth globally with an 89-point improvement over previous benchmarks.

Beyond raw performance metrics, Google emphasizes Carbon's approach to handling complexity and uncertainty. The model seeks human input in ambiguous situations rather than making unsupported claims, reduces unsupported assertions by approximately 50% compared to Argon, and achieves a 0.0% attack success rate on the Gray Swan Indirect Prompt Injection benchmark. Internal testing shows the model reduces task overreach by 26% and under-service by 29% in multi-step workflows. Google also highlights that Carbon demonstrates extended autonomous execution capabilities, with internal observations of the model running continuously for multiple days to solve complex engineering problems across large codebases.

The rollout strategy mirrors Argon's phased approach. Carbon first becomes available to cybersecurity defenders through Google's Fairwind Program this week, with broader access planned for developers, businesses, and consumers. Google is also participating in the U.S. government's voluntary pre-release model testing process. Introductory pricing is set at $2 per 1 million input tokens and $10 per 1 million output tokens, with cached input tokens discounted by 95%. Standard pricing will rise to $4 per 1 million input tokens and $20 per 1 million output tokens after the introductory period.

Industry context shows intensifying competition in frontier AI models. Anthropic's internal research model recently demonstrated the ability to discover critical vulnerabilities in widely used software, including Chrome, Safari, Android, and Windows. OpenAI separately released GPT-6.1 Terra, described as a fast, efficient model for communication and computer-based tasks that meets the Critical threshold in cybersecurity. While Carbon leads several benchmarks, some areas show mixed results—Opus 5.5 maintains advantages in certain software engineering metrics like FrontierSWE 2 and OSWorld-2.1 offline benchmarks.

Topics

Technology Tech Breakthrough Business Science Artificial Intelligence

#Gemini 4 Carbon #vulnerability detection #artificial intelligence #frontier model #prompt injection #machine learning #Google AI #code generation #autonomous agents #cybersecurity

Entities

Why This Matters

Google's Gemini 4 Carbon ranks first across measured AI benchmarks (Artificial Analysis Intelligence Index v5.0: 59 points) and achieves 88.2% accuracy on SEC-bench Pro vulnerability detection, 54.8% on exploitation validation (ExploitGym), and 71.0% on patching (CWE-bench). This affects cybersecurity tool developers and enterprises adopting frontier models for code review and patch generation: the model's two-fold improvement in secure code generation and 10x improvement in security patch quality (per Chrome team) measurably change the capability baseline for automated security workflows. Introductory pricing ($2 per 1M input tokens vs. future $4 standard) creates a time-limited cost structure for early adoption. The phased rollout through the Fairwind Program (cybersecurity defenders first) followed by broader developer/enterprise/consumer access shapes market access timelines. Concurrent competition from Anthropic's internal research model and OpenAI's GPT-6.1 Terra—both demonstrating real-world vulnerability discovery—intensifies the frontier model capability race.

Timeline & Sources

Oct 14, 2026

Wire

Google announced Gemini 4 Argon frontier AI model

Oct 28, 2026

Wire

Sundar Pichai and Google announced Gemini 4 Carbon on social media

Oct 28, 2026

Wire

Carbon begins rolling out to cybersecurity defenders through Fairwind Program

Oct 28, 2026

Wire

Google announced Gemini 4 Carbon frontier AI model via blog post