Google keeps its security model behind a closed door

On September 2, 2026 Google announced two models: Gemini 3.8 Flash for general use, and Gemini 3.8 Flash Cyber, specialized in vulnerability detection and automated patching — available only to "trusted defenders" through a new Fairwind Program.

The gate is the story. Google states plainly that 3.8 Flash ships with safeguards against misuse in CBRN and cyber-offense domains, while 3.8 Flash Cyber ships with a more permissive set of mitigations for cybersecurity, and is therefore restricted to defenders who need the fuller capability set. The restriction is a consequence of the loosened guardrails, not a marketing tier.

What it can do, per the announcement:

• Frontier-level performance on CyberGym, the standard benchmark for autonomous vulnerability discovery, surpassing significantly larger models.
• Over 70% success on an internal benchmark spanning 20 programming languages, built because CyberGym is mostly C/C++.
• On an external patching benchmark run by Collinear, pass@1 of 47.2% against a leading frontier model's 47.8%, at significantly lower cost.
• In production at Google: Chrome Security reports 2.6x more correct patches than much larger commercial models; Wiz reports 7.5–9.7% higher recall on an internal pentesting benchmark at 2.3–5.2x lower cost; the Cloud Vulnerability Research team found a critical foundational vulnerability in under two hours, for something that normally takes months.

Google also says it prioritized vulnerability fixing over offensive capability from the start. That is the trade being made here: a model good enough at finding holes that handing it to everyone is the risk.

Source: Google, "Introducing Gemini 3.8 Flash and 3.8 Flash Cyber" (September 2, 2026).