Google Releases Gemini 3.8 Flash for Agents and Cybersecurity

Google has unveiled two new versions of its Gemini 3.8 Flash model, aimed at enhancing AI capabilities in software development and cybersecurity.

The new models, Gemini 3.8 Flash and Gemini 3.8 Flash Cyber, represent significant advancements over their predecessors. The standard 3.8 Flash is optimized for complex agentic tasks, software engineering, and multi-step reasoning. Google reports that the model has outperformed many large AI models in the DeepSWE coding benchmark at a lower cost.

The Flash Cyber version is described by Google as its most capable cybersecurity model, designed for efficient vulnerability discovery and remediation. The model achieved high scores on the CyberGym and CWE-Bench tests, and in internal evaluations, it identified vulnerabilities across 20 programming languages with over a 70% success rate.

The 3.8 Flash model is available through Gemini Enterprise and for developers via the Gemini API. It is capable of processing large amounts of data and utilizing various formats, including text, images, audio, and video. Google is already using Flash Cyber to secure its own code, and it has demonstrated a higher rate of patching Chrome vulnerabilities compared to larger commercial models.