Introducing Gemini 3.8 Flash and 3.8 Flash Cyber

What it is

Gemini 3.8 Flash Cyber is a gated cybersecurity variant of Google DeepMind's Gemini 3.8 Flash, restricted to "trusted defenders" through the new Fairwind Program. The summary covers the benchmark story; the actionable gap it leaves is who can get this and on what terms — which is the only decision a reader outside Google can act on.

Preconditions / Affected

Fairwind access is not open. The page names three eligible categories:

• Government authorities (trusted)
• Critical infrastructure operators
• Software maintainers

These get "prioritized access" and must "apply for access." The base Gemini 3.8 Flash (non-Cyber) is generally available — Gemini API via AI Studio, Antigravity, Gemini Enterprise, and consumer apps for AI Pro/Ultra subscribers — at $0.75 per million input tokens and $3.75 per million output tokens (same as 3.7 Flash).

How it works (the design stance)

The Cyber variant ships with "a more permissive set of mitigations for cybersecurity" than the general model, which is the stated reason it is gated. Google explicitly states it prioritized vulnerability fixing over offensive capabilities like exploitation from the start. The general 3.8 Flash retains CBRN and cyber-offense misuse safeguards per Google's Frontier Safety Framework; the Cyber model relaxes the cyber ones for vetted users only.

What to do

• If you fall in one of the three eligible categories and want autonomous vuln-discovery/patching, apply to the Fairwind Program (link on the page).
• If you do not, you cannot access the Cyber model — the general 3.8 Flash is what you get, and it is a coding/agent model, not the security-tuned one.

Additional metrics not in the summary

• Wiz reported +7.5–9.7% higher recall on their internal penetration-testing benchmark at 2.3–5.2x lower cost versus other frontier models (the summary states only "higher recall at lower cost").
• Google claims a "significant leap in prompt injection robustness" as measured by Gray Swan — relevant if you're evaluating the model for agentic pipelines exposed to untrusted input.

Caveats

This is a vendor announcement. The 20-language ">70%" and CyberGym results are against an internal, undisclosed benchmark and comparisons to unnamed "significantly larger frontier models"; none of the Chrome/Wiz/Cloud-VRT figures are independently reproducible. The CWE-Bench patching result (47.2% vs 47.8%) is the only externally-run benchmark cited, and it is a near-tie framed favorably on cost, not a capability lead.