Gemini 4 Carbon: capabilities and initial cybersecurity access
Google introduces Carbon with initial access limited to defenders and trusted partners. What it announced and how to read the benchmarks.
Checked against Google’s announcement; results are attributed to the provider.
Google announced Gemini 4 Carbon on 28 October 2026, describing it as its most capable model for cybersecurity and long-horizon tasks. Initial access is limited to a subset of defenders; no firm date for general availability was announced.
What the benchmarks show
Google reports 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench. These tests measure different tasks, including vulnerability discovery, exploitation and patching. They are provider-reported figures, not a substitute for evaluation on your own cases.
Access through Fairwind
Google says it will gradually expand access to trusted partners through the Fairwind Program while gathering feedback and refining safeguards. Availability and conditions may change; check them directly before planning an integration.
How to evaluate it
Define authorised tasks in an isolated test environment. Review code changes, measure false positives and retain human oversight: a benchmark score cannot guarantee that a model finds or fixes every issue.