Thought Leadership
According to Google, the model leads several cybersecurity benchmarks and is initially only available to selected defenders.
Google has presented its new frontier model Gemini 4 Carbon. It is said to bring significant progress, especially in cybersecurity and in tasks over long periods of time. The release follows less than a week after the introduction of Gemini 4 Argon.
“It shows significant progress in programming and agentic tasks as well as in core capabilities for cyber defense, achieving a clear lead on 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
Top of the Intelligence Index
In the Intelligence Index of the independent analysis firm Artificial Analysis, Gemini 4 Carbon achieves 59 points, placing it just ahead of Claude Opus 5.5 from Anthropic with 58 points, as well as GPT-6 Astra with 57 and GPT-6.1 Sol with 55 points from OpenAI. According to Google, Carbon generates secure code twice as often as Argon. In addition, for complex tasks or tasks under pressure, the model is said to ask for clarification when things are unclear, instead of guessing or forcing a path. In Google’s tests, it resisted all evaluated attacks involving indirect prompt injection. Independent tests on this are not yet available.
Access initially for defenders
As with earlier models, Google is initially limiting access to a selection of cyber defense organizations. It is subsequently to be expanded to other trusted partners via the Fairwind program. According to its own statements, Google is also participating in the US government’s voluntary process in which models are tested prior to their release.
Other models with cyber capabilities
In parallel, OpenAI has released the model GPT-6.1 Terra, which is designed for communication and tasks on the computer. According to OpenAI’s evaluation framework, it reaches the “critical” level for cybersecurity capabilities and retains the safeguards developed for GPT-6 Astra. Anthropic had reported in the same week that an internal, unreleased research model found critical vulnerabilities in widespread software. Among them were one-click code execution in Chrome, an exploit chain in Safari, a lock screen bypass on Android, and a denial-of-service vulnerability in the Windows driver HTTP.sys (CVE-2026-33096). Google, Apple, and Microsoft have since patched the vulnerabilities.
(Editorial Team)
