Google this week began offering governments, critical-infrastructure operators and vetted corporate partners access to its most capable cybersecurity-focused AI model, Gemini 3.8 Flash Cyber, through a new gated program called Fairwind — a deliberate choice to restrict rather than broadly release a system the company says can find and patch software vulnerabilities largely on its own.
"Today, we're launching our Fairwind Program, a limited access program for governments and trusted partners to use our most advanced cyber defense capabilities," Four Flynn, Google's vice president of security and privacy, said in the company's announcement. The program prioritizes national cyber authorities, operators of critical infrastructure in healthcare, telecommunications, energy and finance, and core technology platforms; organizations that join must agree to restrict use to internal security, incident-response or penetration-testing teams and to deploy safeguards such as multi-factor authentication.
Google says Flash Cyber exceeds a 70% real-world vulnerability discovery rate and sits at the leading edge of results on CWE-Bench, an industry benchmark for automated patching. The company's Chrome security team found the model produced correct vulnerability patches at 2.6 times the rate of comparable commercial models, according to Tech Times. More than 650 partners, including CrowdStrike, Datadog, Palo Alto Networks and Snowflake, are already participating.
Rivals draw the same line
Google is not alone in choosing to gate rather than broadly ship its most capable offensive-adjacent cyber tools. Anthropic has kept its Claude Mythos 5.1 model restricted to trusted-access programs for cybersecurity and life-sciences work, while allowing a separate model, Claude Fable 5.1, to identify vulnerabilities but redirect penetration-testing and exploit-generation requests elsewhere, according to The Hacker News. OpenAI's forthcoming Astra model, which the company says meets a "Critical" capability threshold for cybersecurity, is being distributed through a similarly gated Daybreak Blue program rather than general release.
The industry-wide pattern reflects a dual-use problem that has grown more acute as models get better at autonomously finding and exploiting software flaws: a system capable of discovering and patching vulnerabilities at scale is, by the same token, capable of discovering and weaponizing them. Google's approach — a fast, comparatively cheap model paired with a mature agent framework, deployed only to vetted defenders — is a bet that controlled access will produce safer real-world outcomes than opening the tool to everyone at once.
Google has not published a formal application process or eligibility checklist for Fairwind beyond describing its priority categories, and the company has not said when, or whether, access might eventually widen. For now, the model's availability will be determined largely by Google's own vetting of each applicant, a gatekeeping role that mirrors the trusted-access structures Anthropic and OpenAI have each built around their own most capable security tools.
