RC RANDOM CHAOS

Gemini 4 Carbon: Google's Cybersecurity Frontier Model

· via Hacker News

Original source

Gemini 4 Carbon

Hacker News →

Google announced Gemini 4 Carbon, its latest frontier model designed to excel at cybersecurity defense and extended reasoning tasks. The model ranks first on Artificial Analysis Intelligence Index v5.0 and demonstrates state-of-the-art performance across the cybersecurity lifecycle: 88.2% on SEC-bench Pro for vulnerability detection, 54.8% on ExploitGym for exploitation validation, and 71.0% on CWE-bench for patching. Internal testing shows a 2x improvement in secure code generation over its predecessor Argon. Chrome’s team reported 10x better performance than another lab’s large frontier model when generating high-quality patches.

Carbon delivers multiple safety improvements. The model achieved 0.0% successful attacks on Terminal-Bench 4.0 with cheating attempts dropping to 0.4%, dropped high-severity misalignment alerts by nearly 80% in live coding environments, and reduced unsupported claims by approximately 2x compared to Argon. On prompt injection resistance, Carbon achieves 0.0% attack success rate on Gray Swan’s Indirect Prompt Injection Benchmark versus Argon’s 0.7%. The model also reduced overreach by 26% and under-service by 29% while improving multi-step task execution.

Google is rolling out Carbon gradually, beginning this week with a limited set of cyber defenders through its Fairwind Program while participating in U.S. government pre-release testing. The model will be available at $2 per million input tokens and $10 per million output tokens, matching Argon’s introductory pricing. Full availability to developers, businesses, and consumers is planned for the coming months.

Read the full article

Continue reading at Hacker News →

This is an AI-generated summary. Read the original for the full story.