Google's Fairwind Program explained: who gets the cyber AI models first and on what terms

Fairwind is the name Google DeepMind gives to its early-access route for security-focused AI models. When Google released Gemini 4 Carbon on October 28, 2026, it went first to a subset of cyber defenders through Fairwind, and the public does not have it. Understanding that program explains why the newest, most capable model is the one you cannot use.

What Fairwind is

Google DeepMind's program page describes Fairwind as giving vetted defenders early access to advanced AI models for cybersecurity. Its tagline is "Keeping defenders one step ahead," and the page says the program "gives trusted partners a critical head start against cyber threats." It says Google works with over 650 partners globally.

The tools on offer are Gemini 4 Argon, which the page describes as a model with cybersecurity defense capabilities for vulnerability detection and automated patching, and CodeMender, a code security agent that can use it. Organizations that are not eligible can still use CodeMender with publicly available models.

Who qualifies

Priority goes to three groups: governments and national cyber authorities, critical infrastructure operators such as healthcare, telecom, energy and financial networks, and core technology platforms. Academic labs focused on defensive benchmarking may also apply. Applicants are vetted, including with background checks, and apply through their Google Cloud account team. Customers who are not on Google Cloud are directed to Google Cloud's get-started documentation.

The terms of access

The conditions are the substance of the program. Partners must use user-level authentication and phishing-resistant multi-factor authentication, and apply access controls. They must limit access to internal cybersecurity, incident-response or penetration-testing teams, and track employee access and use. They may not share, redistribute or resell access. Only permitted dual-use tasks are allowed, such as authorized threat simulation, reverse engineering and malware analysis for defensive or academic research.

These are the kinds of conditions a security vendor would expect. They do not describe a monitoring or audit regime beyond the partner tracking its own use, and the program page does not say what happens if a partner breaks the rules.

How a model moves through it

Google's own posts show the pattern. For Gemini 4 Argon, Google's October 13 announcement says the model is available without cyber guardrails to members of the Fairwind Program, which is expanding by hundreds of partners, and also broadly to developers through its API and to Google AI Ultra and Pro subscribers. WinBuzzer reports Argon was first announced for Fairwind partners on September 30, so the gap between gated and public was about two weeks.

Carbon follows the same first step. Google's Carbon post says it is available starting the week of announcement to a subset of cyber defenders and will expand gradually to more trusted partners through Fairwind. Broader availability to developers, businesses and consumers has no date. The post says "as soon as possible." Carbon's cyber scores, including 88.2% on SEC-bench Pro and 54.8% on ExploitGym, are the reason the first audience is defenders, and we examine them in our report on Gemini 4 Carbon.

Why gate at all

A model good at finding and exploiting vulnerabilities helps attackers as much as defenders. Gating gives defenders a window to patch before the same capability reaches everyone. Google also says it is participating in the US government's voluntary pre-release model testing process for Carbon. The weakness is that the benefit depends on the window being long enough and on partners keeping to the terms. The two-week gap for Argon is short against the time many organizations take to patch.

What the program page leaves out

The page does not give a launch date for Fairwind. It does not say how many partners have received Carbon, how long vetting takes, or what Google does when a partner breaks the terms. It names two products, Argon and CodeMender, and Carbon appears in Google's launch post rather than on the program page we read.

Google's Argon post gives a few facts about early use. It says Wiz, which runs the Scan for Good initiative, used early access to report results in government and healthcare security work. It also describes the safeguards for the public version: external bio red teamers could unlock less than 1% of the bio sub-tasks tested, and shortcut attempts on coding benchmarks dropped to 2%. Those are Google's own figures, and they describe the public Argon, whereas Fairwind members get it without cyber guardrails. The public version and the gated version are therefore different products, which is the point of the program and the reason outside tests of one say little about the other.

What to watch

Three markers will tell you whether gating is working. The first is the gap between Fairwind and public release for Carbon compared with Argon. The second is whether Google publishes results from the government pre-release testing. The third is whether other labs adopt comparable programs; our tracker of pre-release AI model testing follows who tests what.

Applicants should also expect the process to take time, since vetting includes background checks, though the page gives no timeline.

Our take

Fairwind is a reasonable answer to a real problem, and its terms are more concrete than most release policies. But it is a company-run program with no published audit, and the only public evidence of its effect is that Argon reached everyone after two weeks. Judge it by the next release: a longer gap, a published result and a named external check would make the head start real.