Google has released two new AI models this week. The base model is called Gemini 3.8 Flash. Alongside it comes Gemini 3.8 Flash Cyber. The second model targets automated vulnerability discovery directly. It is built specifically for verified security defenders.
Gemini 3.8 Flash brings major upgrades over its predecessor model. It improves noticeably at coding and reasoning tasks. It also improves at long, multi-step problem solving.
Google says it approaches costlier frontier models now. Pricing stays at $0.75 per million input tokens. Output tokens cost $3.75 per million, per the Google Blog announcement.
A Model Built for Autonomous Agents
The new model targets agentic workflows quite directly. It handles long-horizon software engineering tasks reliably. It also performs well in specialized professional fields.
These include finance work and legal analysis tasks. The model reportedly “works harder” on complex problems. It runs extra reasoning steps completely on its own.
Developers can access 3.8 Flash through enterprise deployments now. It is also live inside Google AI Studio. Consumers get access through the Gemini app and Google Search. Enterprise rollouts already reach thousands of on-site Gemini engineers.
Flash Cyber Hunts Software Vulnerabilities
Gemini 3.8 Flash Cyber serves a narrower defensive purpose. It finds security flaws inside real software systems.
On Google’s internal benchmark, it succeeds over 70 percent of attempts. That benchmark spans twenty different programming languages.
On CWE-Bench, the model scores 47.2 percent pass rate. That result nearly matches leading frontier security models today. The model also handles automated patching work well.
Google prioritized defensive fixes over offensive exploit capability, the company stated clearly.
Restricted Access Through the Fairwind Program
Flash Cyber is not open to everyone yet. Google built a new gatekeeping system for it. The system is officially called the Fairwind Program. It limits access strictly to verified security defenders.
This mirrors growing caution around AI use across cybersecurity generally. Google wants to prevent dangerous offensive weaponization outright. Other AI vendors face similar scrutiny these days. OpenAI’s Astra model was recently rated a critical risk.
Why This Release Matters
AI models increasingly secure real-world infrastructure now. Google’s own Chrome team already uses Flash Cyber. It produced 2.6 times more correct patches. That beat much larger commercial models, VentureBeat reported.
Security firm Wiz also tested the model internally. It found notably higher vulnerability recall than rivals. The tool worked at a far lower cost. This signals a broader shift toward specialized defense AI.
Enterprises already rely on Gemini-powered products across Google’s ecosystem. Adding autonomous cyber-defense tools extends that footprint even further. Expect rivals to respond with competing security-focused models soon.
