Google's Gemini 3.5 Flash Unveils New Vulnerability Hunting Capabilities

Post Views: 13

Google’s Gemini 3.5 Flash Cyber model identifies, verifies, and addresses security flaws prior to exploitation while reducing potential misuse risks.

CodeMender Overview

CodeMender is an AI-driven code analysis system developed by Google DeepMind, part of a restricted pilot initiative expanding to governmental entities and verified collaborators via CodeMender. It functions as a managed code security agent, offering users direct access to scanning and remediation features in a preview phase.

Key Features

CodeMender provides entry to widely available models through the Gemini Enterprise Agent Platform or can be implemented as a foundational element of AI Threat Defense, according to Michael Gerstenhaber, VP of Product Management for Gemini Enterprise, and Clemens Viernickel, Director of Product Management for Cloud AI.

Efficient code security analysis requires examining extensive execution paths within software. Traditional large language models often hinder scalability and slow down processes. Gemini 3.5 Flash Cyber is engineered to detect flaws across expansive codebases by evaluating multiple execution paths.

How Gemini 3.5 Flash Cyber Works

CodeMender leverages the model iteratively to analyze these paths, identify and confirm vulnerabilities, and compile reports through sub-agents. The system integrates with regular security audits, time-sensitive release cycles, and commit-scanning workflows at scale.

Testing and Results

Google tested Gemini 3.5 Flash Cyber using CyberGym, a framework developed by its Big Sleep team, and within Chrome’s production commit-scanning pipeline. The model demonstrated performance comparable to larger models on CyberGym and surpassed Gemini 3.5 Flash in evaluations conducted by the Big Sleep team and Chrome’s pipeline.

During trials on the V8 JavaScript engine, the model uncovered 55 distinct confirmed issues, including 10 that were undetected by Gemini 3.5 Flash and Claude Opus 4.6. Google employs the tool to address vulnerabilities across internal systems, such as Chrome, Android, Cloud, Ads, and YouTube.

Expert Insights

Raluca Ada Popa, Head of Security Privacy Research at Google DeepMind, and Four Flynn, VP of Security and Privacy at Google DeepMind, noted that the model identified remote-code-execution vulnerabilities in public APIs and a memory-corruption flaw in a critical production service within two hours. It also produced a fully reliable remote-code-execution exploit capable of bypassing Address Space Layout Randomization and Write XOR Execute protections.

New Gemini Models

Building on Gemini 3.5 Flash, Google launched two new models: Gemini 3.5 Flash-Lite and Gemini 3.6 Flash. Gemini 3.5 Flash-Lite targets low-latency, high-throughput tasks like agentic search and document processing. It supports adjustable reasoning levels and includes integrated computer-use capabilities for agentic workflows.

Enhancements in Gemini 3.6 Flash

Gemini 3.6 Flash is optimized for coding, knowledge-based tasks, multimodal operations, and agentic workflows. It reduces output tokens, reasoning steps, and tool calls compared to Gemini 3.5 Flash, lowering costs for multi-step processes while enhancing performance in coding, computer use, and knowledge benchmarks.

“3.6 Flash incorporates enhanced Frontier Safety measures in Chemical, Biological, Radiological, and Nuclear (CBRN) and cyber offense misuse domains. These safeguards significantly increase resistance to jailbreak attempts while minimizing refusals for beneficial applications,” stated Tulsee Doshi, Senior Director of Product Management, on behalf of the Gemini team.

Additional Topics

AI code analysis cybersecurity Agentic AI developments Google’s advancements in AI coding agents Ransomware targeting AI infrastructure Zero-day exploits in SonicWall systems