Gemini 3.8 Flash Cyber — détection/patching réservés aux défenseurs

Blog Google 2026-09-02. Détection/patching défensifs gated Fairwind. Benches vendor ; FLAG % Wiz. Pas de how-to exploit.

Defensive framing only. Google announced Gemini 3.8 Flash Cyber on 2 September 2026 as a cybersecurity-oriented model available to trusted defenders via the Fairwind Program, paired with CodeMender-style harnessing for finding and validating patches inside an organization’s secure environment.

Vendor-stated benches: Google describes frontier-level CyberGym performance vs prior Cyber / larger models; an internal multi-language vuln benchmark with success >70% across 20 languages; and Chrome Security noting 2.6× more correct patches vs best commercial models in their test. FLAG · third-party: Google cites Wiz reporting about +7.5–9.7% higher recall on Wiz’s internal pentest benchmark at lower cost — treat as third-party / partner-reported %, not an independent public audit. Pricing for the Cyber deployment is characterized as Flash-level cost by the vendor; do not invent dollar rates when no public Cyber rate card is cited.

For MSPs: map Fairwind-style access to customer change control — detection and patch proposals still need human review, ticket linkage, and deploy gates. This article does not describe exploit development or offensive tradecraft.

FAQ

Qui peut utiliser Flash Cyber ?

Google le positionne pour des défenseurs de confiance via Fairwind. Le Flash public est un déploiement général séparé.

Comment traiter le pourcentage Wiz ?

Comme métrique FLAG third-party / partenaire sur le benchmark interne Wiz — contexte, pas un substitut à vos evals défensives.

Inference EU-facing et isolation tenant restent des questions produit séparées.

Sources

Draft only — do not publish without editorial review.