Google DeepMind Launches Gemini 3.8 Flash with Restricted Cyber Version

Google DeepMind released Gemini 3.8 Flash on September 3, 2026, along with a restricted version called Gemini 3.8 Flash Cyber, designed specifically for government and critical infrastructure security teams.

The Flash Cyber version has been classified as having "critical" cyber capabilities, meaning it can identify and exploit software vulnerabilities with minimal human oversight. Google is restricting access to vetted organizations and critical infrastructure operators to prevent misuse.

The release follows a similar move by OpenAI, whose forthcoming "Astra" model has also been identified as having critical cyber capabilities. Both companies are departing from the open-access model that characterized earlier generations of AI tools.

Security researchers say the classification of AI models by their potential for harm represents a significant shift in how the industry approaches frontier capabilities. The concern is that models capable of autonomous vulnerability exploitation could be used by malicious actors if released without controls.

The standard Gemini 3.8 Flash model is available more broadly and is optimized for speed and efficiency in everyday tasks. Google says it performs well on coding, reasoning, and multimodal tasks.

The G20 recently endorsed the U.S.-backed "Carolina Principles," a set of non-binding, sector-specific guidelines for AI regulation that favor targeted oversight over sweeping international laws. The principles are expected to influence how governments approach the regulation of high-capability AI models like Flash Cyber.

Google has not disclosed the full list of organizations that will have access to the Flash Cyber version, citing security concerns.