Google launches Fairwind cyber program with more than 650 partners

Google launched the Fairwind Program on September 2, a limited access program that hands its cyber defence tools to vetted governments and companies instead of to everyone. The company says more than 650 partners are taking part globally.

The offering pairs Gemini 3.8 Flash Cyber, which Google calls its most advanced cyber model, with a harness called CodeMender. Together they’re meant to help defenders find, verify and fix vulnerabilities at what Google describes as agentic scale. The announcement is bylined to Four Flynn, vice president for security and privacy.

Google’s argument is that a defender’s edge comes from shrinking the time between detecting a flaw and patching it, so the pitch here is speed. Google says teams that used to spend weeks fixing a flaw by hand can generate verified, deployment-ready patches in minutes, inside their own secure cloud environment. It also claims the pairing runs at a fraction of the operating cost of traditional frontier models, though it publishes no benchmark figures or pricing to check that against.

Spotting weaknesses creates awareness and fear; autonomously finding and fixing vulnerabilities delivers security.

Four Flynn, VP Security and Privacy, Google

Access comes with conditions. Participating organisations agree to limit use to employees on their internal cybersecurity, incident response or penetration testing teams, and to deploy protections such as multi-factor authentication. Google is staging entry by category: governments and national cyber authorities, critical infrastructure operators, and core technology platforms.

That’s roughly the trade every lab has landed on this year, so Fairwind is a late entry rather than a first. OpenAI runs Trusted Access for Cyber, and AWS opened its Daybreak Red and Daybreak Blue tiers to eligible Bedrock customers on August 11, in one Region, US East (Ohio). Anthropic runs its own Cyber Verification Program, according to TechCrunch. We covered the Bedrock rollout when it happened.

Program	Company	What it gates	Stated purpose
Fairwind	Google	Gemini 3.8 Flash Cyber with CodeMender	Finding and fixing vulnerabilities
Daybreak Blue	OpenAI	GPT-5.6 Sol	Vulnerability discovery, detection engineering, incident response
Daybreak Red	OpenAI	GPT-5.6 Cyber	Vulnerability research, exploit reproduction, mitigation development
Cyber Verification Program	Anthropic	Not detailed in these sources	Not detailed in these sources

Vetting has a failure mode, and it surfaced in August. TechCrunch reported on August 19 that several researchers lost their Trusted Access for Cyber privileges, and OpenAI said an error caused it. All five researchers the outlet spoke to live outside the United States and Europe, which hints the revocations landed unevenly by region. Anyone applying to Fairwind is accepting the same dependency on one vendor’s verification pipeline.

Google’s own page leaves gaps too. The partner testimonials are published as images, credited in their alt text to CrowdStrike, Wiz, Snowflake, Palo Alto and Armadin, but the partner roster is a logo graphic rather than a list, which means the 650 figure isn’t something you can audit from the post, and no application timeline is given.

There’s a route in for everyone else. Google says any Google Cloud customer can use CodeMender with publicly available models hosted on Gemini Enterprise Agent Platform, alongside its AI Threat Defense products. Separately, Google.org put its total cybersecurity funding above $100 million globally, including $36 million for 35 cyber clinics that have supported over 1,250 hospitals, public school districts and municipal utilities in the US.

Watch the security vendors Google names as partners, because several are building the same capability themselves. CrowdStrike, whose testimonial sits on the Fairwind page, used its Fal.Con conference on the same day to describe a cyber superintelligence lab. Its models showed a 29% higher detection rate and remediated threats six times faster than leading alternatives in the company’s own evaluations, chief AI and autonomous systems officer Bartley Richardson told theCUBE.

Those are vendor numbers, not independent tests, and that’s the open question under all of this. Nobody outside the programs can measure whether or not gated models close flaws faster than attackers find them. The next thing worth reading is a partner disclosure with patch counts in it, not another launch post. Our earlier piece on approval-only cyber models tracks how narrow this door has become.