Google on Wednesday took the wraps off Gemini 4 Carbon, its latest frontier model with state-of-the-art cybersecurity and long-horizon capabilities.
“It shows significant gains across coding and agentic tasks as well as across key capabilities critical for cybersecurity defense, landing a firm lead in SEC-bench Pro (88.2%) for finding vulnerabilities, ExploitGym (54.8%) for exploitation, and CWE-bench (71.0%) for patching,” Koray Kavukcuoglu, senior vice president of Google DeepMind and chief AI architect at Google, said.
On the Artificial Analysis Intelligence Index benchmark, Carbon leads with a score of 59, surpassing Anthropic’s Claude Opus 5.5 (58), OpenAI GPT-6 Astra (57), GPT-6.1 Sol (55).
One of the core focus areas is cybersecurity, where the latest iteration reportedly delivers significant performance improvements across the board, including vulnerability detection, validation, and remediation. Carbon is also twice as likely to generate secure code versus Argon.
Google also emphasized that Carbon attempts to respond to complex or high-pressure tasks by seeking human input in the face of ambiguity rather than “guessing or forcing a path forward,” as well as resisting every evaluated attack attempt at indirect prompt injection.
The debut of Gemini 4 Carbon comes less than a week after the tech giant unveiled Gemini 4 Argon. The release plan is similar to that of its predecessor, with the rollout currently limited to a subset of cyber defenders before expanding access to other trusted partners as part of its Fairwind Program.
Google also said it’s actively engaged in the U.S. government’s voluntary process for pre-release model testing.
The development comes as OpenAI released GPT-6.1 Terra, which it described as a fast, efficient model that “excels at communication and computer-based tasks.”
“GPT‑6.1 Terra meets the Critical threshold in cybersecurity under our Preparedness Framework, and maintains the protections developed for GPT‑6 Astra,” OpenAI said.
“GPT-6.1 Terra brings with it the alignment improvements of the GPT-6 series. According to our evaluations, GPT-6.1 Terra is more likely to stay within scope and honestly disclose limitations and potential blockers to the user compared to GPT-5.6 Terra. In our prompt-injection tests, GPT-6.1 Terra also showed strong resistance to malicious instructions designed to divert it from the user’s task.”
Earlier this week, rival Anthropic also disclosed that an unreleased, internal research model uncovered critical vulnerabilities in some of the world’s most widely used software.
This included a one-click remote code execution in Google Chrome, a one-click exploit chain in Apple Safari that could be used to access a user’s browsing history, cookies, and other sensitive data, an Android lock screen bypass, and a denial-of-service (DoS) flaw in the Windows HTTP.sys driver (CVE-2026-33096, CVSS score: 7.5). All the identified issues have been fixed by Google, Apple, and Microsoft.
“These attack chains demonstrate the model’s ability to create sophisticated, multi-step exploits in both open-source and proprietary software,” Anthropic said.
“These skills have historically been the domain of few specialist teams with expertise in these targets. Now, it’s something models can do with little human involvement, and then prove by producing a working exploit that anyone can run and verify.”
