Google announced Gemini 4 Carbon, its new and most powerful AI model, but not everyone can use it. According to SecurityWeek and TechRepublic, in the first phase the model is available only to selected cyber defenders, through a program called Fairwind that includes governments, critical infrastructure operators and large tech companies.
Why the restriction
Carbon is especially good at exactly what is dangerous. According to Google's own figures, it scored 88.2% on a vulnerability discovery benchmark, 54.8% on exploitation and 71% on fixing vulnerable code. A model that can find and exploit flaws can help defenders fix them, but also help attackers break in. So Google wants defenders to get it first.
When for everyone
Google plans to widen availability later but has not set a date. Carbon is not yet in the public API or available to consumers, and developers are pointed to Gemini 4 Argon, launched two weeks earlier. All figures are Google's, and according to one report they were tested without the model's guardrails.
An industry trend
Google isn't alone: OpenAI, Anthropic and Mistral have also restricted the cyber capabilities of powerful models, or given them to defenders first. It illustrates the central dilemma of AI in cybersecurity: the same tool that can close thousands of holes before attackers find them can also open them.
