· 10 min read

Google just launched a cybersecurity AI that beats everything else at finding bugs, and you cannot apply for access

AI Tools

Google DeepMind shipped Gemini 3.8 Flash Cyber on September 2, the most capable vulnerability-hunting model it has ever released, scoring 86.2% on the CyberGym benchmark and beating GPT-5.5-Cyber, Mythos 5, and its own predecessor at finding real exploitable bugs. Then it put the model behind a program called Fairwind that, by its own published FAQ, has three eligible categories: governments, critical infrastructure operators, and "core technology platforms" serving millions of downstream users. I run patch work for small WordPress installs and a couple of client SaaS backends. I am none of those three things, and neither is almost anyone reading this.

What actually shipped on September 2

Google announced two models the same day: the general-purpose Gemini 3.8 Flash and the security-specialized Gemini 3.8 Flash Cyber. They share the same underlying training, but Google gave Flash Cyber a more permissive set of safety mitigations so it can do vulnerability research work the standard model blocks or throttles. On CyberGym, the industry-standard benchmark for autonomous vulnerability discovery, Flash Cyber hit 86.2% pass@1, up from 77.5% for the previous Gemini 3.5 Flash Cyber and ahead of GPT-5.5-Cyber's 85.6%. On an internal Google benchmark spanning 20 programming languages, it found real vulnerabilities more than 70% of the time. On CWE-Bench, the external patching benchmark run by Collinear, it landed at 47.2% pass@1, essentially tied with a much larger frontier model's 47.8%, at a fraction of the cost.

Those aren't marketing numbers pulled from a slide. Google's Chrome security team reported 2.6 times more correct patches from Flash Cyber than from larger commercial models they tested against. Wiz measured 7.5 to 9.7 percentage points higher recall on its internal penetration testing benchmark, at 2.3 to 5.2 times lower cost than the frontier models it normally runs. Google's own Cloud Vulnerability Research team says it found a critical foundational vulnerability in under two hours with the model, work that would typically take months. I don't love taking a vendor's internal numbers at face value, but three independent-ish sources (Chrome, Wiz, Cloud VRP) landing in the same range is more corroboration than most AI security claims get.

The Fairwind Program, and who actually qualifies

Flash Cyber isn't sold, and it isn't available through the Gemini API, AI Studio, or anywhere a normal developer account can reach it. It's distributed exclusively through the Fairwind Program, a new access-control layer DeepMind launched alongside the model. According to DeepMind's own FAQ, the eligible categories are: governments and national cyber authorities defending public-sector networks, critical infrastructure operators protecting healthcare, telecom, energy, and financial networks, and "core technology platforms" whose software underpins millions of downstream users. Google's Fairwind page says it already works with more than 650 partners globally, including named references from Armadin, CrowdStrike, Palo Alto Networks, and Snowflake.

Fairwind partners get Flash Cyber paired with CodeMender, DeepMind's separate patch-generation agent that finds root causes and validates fixes automatically before a human ever reviews them. That combination, find the bug and generate a tested patch in one pass, is the actual product. It's not "here's a smarter chatbot," it's "here's an agent loop that can triage your CVE backlog while you sleep." I want that. I run security patch cycles for clients who pay me specifically because they don't have anyone else to do it, and a tool that catches a root cause in two hours instead of two months would change how I price that work.

Here's where the exclusion gets concrete rather than theoretical: DeepMind's own FAQ answers "can academic institutions apply?" by directing students toward CodeMender on Google Cloud instead of Fairwind, and answers "what can I do if I'm not eligible" by pointing rejected applicants at CodeMender with publicly available models, not at Flash Cyber under any circumstances. A one-person security consultancy patching a client's WordPress install doesn't fit governments, doesn't fit critical infrastructure, and doesn't fit "core technology platform serving millions." Google isn't hiding this. It's just not built for the long tail of the internet, which happens to be where most of the internet's actual attack surface sits: the small business sites, the abandoned plugins, the SaaS backends built by a team of two.

The gap this creates

I want to be fair to Google's logic here, because it isn't arbitrary. Flash Cyber ships with fewer safety mitigations specifically so it can do dual-use work: the same capability that finds a bug for a defender finds it for an attacker. Restricting access to vetted organizations with background checks, phishing-resistant MFA, and tracked employee access is a real mitigation, not security theater. Anthropic did something similar with Claude Mythos and Project Glasswing back in April, and OpenAI gates GPT-5.6-Cyber the same way. This is becoming the standard playbook for frontier cyber models, not a Google-specific overreach.

But the consequence lands unevenly. A bank's internal security team gets a model that finds a critical vulnerability in two hours. A solo operator maintaining forty client WordPress sites, which collectively probably have more real-world exposed vulnerabilities than any single bank's codebase, gets the plain Gemini 3.8 Flash: same underlying intelligence, deliberately less capable at the exact task that matters here. That's not a hypothetical gap, it's the stated design of the product. The tools that would help the most exposed part of the internet the most are reserved for the part of the internet that's already best defended.

What I'd actually do this week

If you're in the same position I am, here's the honest read: don't wait for Fairwind eligibility, because it isn't coming for a solo shop. CodeMender itself is in public preview on the Gemini Enterprise Agent Platform right now, and it works with generally available models, not just the gated Cyber variant. It integrates into VS Code, Antigravity, and a CLI. That's the actual on-ramp DeepMind points rejected applicants toward, and it's worth setting up even paired with the weaker model, because a patch-generation agent that catches root causes is still useful when the underlying vulnerability-finder is merely good instead of best-in-class.

Pair that with the boring stuff that was already true before any of this shipped: run WPScan against client WordPress installs on a schedule, keep Dependabot or Renovate wired into every repo you touch, and don't treat "the client hasn't been breached yet" as evidence of anything. None of that closes the gap between what a 650-partner Fairwind member gets and what I get. It just means I'm not pretending the gap doesn't exist while doing nothing about it.

Where I could be wrong: Google's FAQ says it "routinely reviews applications and adds new partners over time," and 650 partners in a few months of the Fairwind Program (it followed a Gemini 3.5 Flash Cyber pilot that started in July) is a fast ramp. If that keeps accelerating, category three, core technology platforms, could stretch to include software agencies or larger consultancies before this time next year. I wouldn't bet a business plan on it, but the exclusion isn't necessarily permanent, just structural for now.

Author

Lukas

Sources

More AI articles →

Stay in the Loop

Get new posts delivered to your inbox. No spam, unsubscribe anytime.

Newsletter coming soon. Set PUBLIC_CONVERTKIT_FORM_ID in .env to activate.

Related Posts