Google's Gemini 3.5 Flash Cyber AI Hunts Vulnerabilities in Chrome & Safari

Google DeepMind has launched Gemini 3.5 Flash Cyber, a specialized AI model built atop Gemini 3.5 Flash, engineered to autonomously discover, validate, and patch software vulnerabilities at scale. The model is being deployed exclusively through CodeMender, an AI-powered vulnerability discovery and patching agent that Google first unveiled in October 2025. As part of a limited-access pilot program, access is restricted to governments and trusted partners, with DeepMind citing the dual-use nature of the technology as the central reason behind its restricted rollout.

The lightweight model is positioned as a cost-efficient alternative to larger, resource-heavy cybersecurity-focused AI systems. CodeMender can invoke Gemini 3.5 Flash Cyber multiple times at high speed and low cost, allowing the agent to scan far more code paths than competing models would typically permit. "Given the dual-use nature of this technology, we have taken an intentional approach to how we deploy 3.5 Flash Cyber," said Raluca Ada Popa, DeepMind's Gemini Security Lead, and Four Flynn, vice president of security and privacy at DeepMind, in a blog post shared ahead of publication. Because the model runs solely inside CodeMender, DeepMind noted it is straightforward to enforce guardrails that enable defensive functions while disabling offensive cyber activity, preventing scenarios where a model refuses actions critical to AI-assisted forensic analysis.

In internal evaluations, Gemini 3.5 Flash Cyber consistently outperformed Gemini 3.5 Flash, Gemini 3.6 Flash, and Anthropic's Claude Opus 4.6 across complex codebases. Stress-testing on Google Chrome and Apple Safari's V8 JavaScript Engine yielded 55 unique confirmed vulnerabilities within a fixed number of invocations, compared to 47 found by Gemini 3.5 Flash and just 36 by Claude Opus 4.6, including 10 issues the other models missed entirely. The release lands alongside Gemini 3.6 Flash and 3.5 Flash-Lite, which target improved coding, knowledge work, multimodal performance, and low-latency tasks respectively, while DeepMind confirmed plans to extend 3.5 Flash Cyber with red-teaming capabilities and end-to-end enterprise defense.

For defenders monitoring their own exposure surface, complementary hygiene checks remain essential. Running a port scanner and an SSL/TLS checker can surface misconfigurations before adversaries exploit them, while an email breach checker offers a quick way to verify whether developer credentials tied to internal repositories have already surfaced in leaked datasets.