Blog Trust

The AI Model Google Won't Let You Touch: Gemini 3.8 Flash Cyber

What Google released

On September 2, 2026, Google launched two models. Gemini 3.8 Flash is the general one: Google’s most intelligent workhorse model, with gains over 3.7 Flash in software engineering, agentic tasks and multi-step reasoning, at an introductory 0.75 dollars per million input tokens and 3.75 per million output through the end of the year. Then there is its sibling, Gemini 3.8 Flash Cyber, which you almost certainly cannot use.

Watch the video (1:17): The AI Model Google Won't Let You Touch: Gemini 3.8 Flash Cyber

The model you cannot download

Google calls Flash Cyber its most capable cybersecurity model, with frontier-level performance at finding and patching vulnerabilities. Its headline number comes from Google’s own Chrome Security team, which reported that Flash Cyber produced 2.6 times more correct patches to Chrome vulnerabilities than the best commercial models, all of which are much larger. But it ships only through a restricted access program called Fairwind, limited to trusted government authorities, critical infrastructure operators and software maintainers. There is no consumer tier and no open API. You apply, and Google decides.

Why gate it

A model good at writing security patches is, by definition, good at understanding vulnerabilities, and that knowledge cuts both ways. Both 3.8 Flash models carry safeguards against chemical, biological, radiological and nuclear misuse and against cyber-offense. The difference is that Flash Cyber runs a more permissive set of cybersecurity mitigations so it can do real defensive work, and Google offsets that by handing the keys only to vetted defenders. The gate is the safety mechanism, not an afterthought to it.

Why it matters for the stack

This is a preview of how the most capable models may ship from here: not a launch, but an allowlist. When a tool is strong enough to be dual-use, access control becomes part of the product. For most builders, Flash Cyber is a signpost rather than a dependency. It shows that the frontier now includes models you will read about but never run, and that “who is allowed to use it” is becoming as much a spec as speed or price.

Sources: for this explainer we checked Google’s Gemini 3.8 Flash and Flash Cyber announcement.

Common questions

Why can't I access Gemini 3.8 Flash Cyber?
It is limited to Google's Fairwind Program, open only to vetted government authorities, critical infrastructure operators and software maintainers. You apply for access; you cannot just sign up.
How is it different from regular 3.8 Flash?
It is the same Flash generation, tuned for security work and released under a more permissive set of cybersecurity mitigations. Standard 3.8 Flash keeps tighter restrictions on cyber-offense uses.
Is the 2.6x claim independent?
No. The figure comes from Google's own Chrome Security team, which reported 2.6 times more correct Chrome vulnerability patches than much larger commercial models.

Watch on YouTube

New posts, in your inbox