Google Takes the Lead in AI Cybersecurity Benchmarks with Gemini 4 Carbon Release

Google has officially taken the wraps off its latest frontier model, Gemini 4 Carbon, boasting state-of-the-art cybersecurity and long-horizon capabilities designed to transform how developers and security teams handle software vulnerabilities. Unveiled on Wednesday by Koray Kavukcuoglu, senior vice president of Google DeepMind and chief AI architect at Google, the new model arrives hot on the heels of the company’s previous rollout and marks a significant escalation in the ongoing race for artificial intelligence supremacy among tech giants.

According to Google DeepMind, Gemini 4 Carbon demonstrates substantial performance gains across coding and agentic tasks. Most notably, however, the model excels across key capabilities critical for modern cybersecurity defense. Kavukcuoglu highlighted that the model has secured a firm lead in several industry-standard evaluations, landing at 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation tasks, and 71.0% on CWE-bench for patching security flaws.

The competitive landscape for foundational AI models has grown increasingly intense, and Gemini 4 Carbon’s benchmark performance reflects these shifting dynamics. On the Artificial Analysis Intelligence Index benchmark, Carbon currently leads the pack with an impressive score of 59. This performance edges out prominent rival models, including Anthropic’s Claude Opus 5.5, which scored 58, as well as OpenAI’s offerings, such as GPT-6 Astra at 57 and GPT-6.1 Sol at 55.

A core focus area for the development of Gemini 4 Carbon was robust cybersecurity defense. The latest iteration reportedly delivers sweeping performance improvements across vulnerability detection, validation, and remediation workflows. In a notable metric shared by Google, Carbon is twice as likely to generate secure code compared to its predecessor, Argon.

Beyond its technical proficiency in identifying and resolving software flaws, Google emphasized that Carbon incorporates advanced behavioral safeguards. When faced with complex or high-pressure tasks involving ambiguity, the model is designed to seek human input rather than guessing or forcibly carving out a path forward. Furthermore, Google reported that Carbon has successfully resisted every evaluated attack attempt targeting indirect prompt injection, showcasing heightened resilience against sophisticated manipulation techniques.

The debut of Gemini 4 Carbon comes less than a week after the tech giant unveiled Gemini 4 Argon. The release strategy for Carbon mirrors that of its predecessor. Access to the frontier model is currently limited to a carefully selected subset of cyber defenders before Google gradually expands availability to other trusted partners as part of its Fairwind Program. Additionally, Google confirmed that it is actively participating in the United States government’s voluntary process for pre-release model testing, aligning with industry-wide pushes for transparent AI safety practices.

The timing of Google’s announcement coincides with major movements from its primary competitors in the artificial intelligence sector. OpenAI recently released GPT-6.1 Terra, describing it as a fast and efficient model that excels at communication and computer-based operations. According to OpenAI, GPT-6.1 Terra meets the critical threshold in cybersecurity under its Preparedness Framework while maintaining the foundational protections developed for GPT-6 Astra.

OpenAI also noted that GPT-6.1 Terra benefits from the alignment improvements inherent to the GPT-6 product series. Internal evaluations suggest that the new model is more likely to stay within defined operational scopes and honestly disclose its limitations and potential blockers to the user compared to its predecessor, the GPT-5.6 Terra. Furthermore, OpenAI’s prompt-injection tests indicated that GPT-6.1 Terra demonstrates strong resistance against malicious instructions designed to divert the system away from the user’s primary objectives.

Meanwhile, the broader discussion surrounding AI and cybersecurity was recently amplified by rival firm Anthropic. Earlier in the week, Anthropic disclosed that an unreleased internal research model successfully uncovered critical, zero-day vulnerabilities in some of the world’s most widely used software packages.

The discoveries made by Anthropic’s research model included a one-click remote code execution flaw in Google Chrome, a sophisticated exploit chain in Apple Safari capable of harvesting a user’s browsing history, cookies, and other sensitive data, an Android lock screen bypass, and a denial-of-service vulnerability in the Windows HTTP.sys driver, tracked as CVE-2026-33096 with a CVSS score of 7.5. Google, Apple, and Microsoft have since issued patches to resolve all of the identified issues.

Anthropic pointed out that these sophisticated attack chains illustrate the model’s emerging capability to autonomously create multi-step exploits targeting both open-source and proprietary software environments. Historically, such specialized skills were restricted to a small number of elite security teams possessing deep domain expertise. The rapid advancement of artificial intelligence, however, means these complex offensive capabilities are increasingly accessible to models operating with minimal human involvement, capable of proving their findings by producing working exploits that can be readily executed and verified by external observers.