Google releases Gemini 4 Carbon to cyber defenders first under a staged rollout

A release designed to reach defenders first

The Fairwind Program covers more than 650 partners, including governments and national cyber authorities, critical-infrastructure operators in healthcare, telecommunications, energy and finance, and core technology platforms, under conditions such as background checks, phishing-resistant multi-factor authentication, use restricted to internal security teams, and no redistribution. Malicious tasks such as creating malware are not permitted. Google said it is taking part in the US government's voluntary pre-release model testing process. After Argon, which went to Fairwind partners on 30 September and to general availability on 13 October, defender-first release followed by broader access is becoming the company's standard sequence.

Published scores and an unpublished safety framework

The figures Google says it measured without guardrails are 88.2% on SEC-bench Pro, 54.8% on ExploitGym (Argon: 42.0%) and 71.0% single-attempt on Collinear CWE-bench. On safety it cites a 0.0% successful cheating rate on Terminal-Bench 4.0, roughly 80% fewer high-severity misalignment alerts internally, and a 0.0% attack success rate on the Gray Swan indirect prompt injection test (Argon: 0.7%).

Independently, Artificial Analysis ranks the model first on its index (59) while its measured hallucination rate rose from Argon 4% to 11%. The reporting available carries no model card and no determination under the Frontier Safety Framework critical capability levels. Offensive capability scores are thus public while the level at which that capability has been judged remains unpublished.