Gemini 3.8 Flash Cyber Finds Vulnerabilities and Produces 2.6x More Valid Fixes in Chrome Testing

Google DeepMind Sep 3, 2026

Key Info

Google DeepMind says the new Gemini 3.8 Flash Cyber model excels at autonomously finding weaknesses on security benchmarks like CyberGym while remaining fast and efficient. In real-world testing on Google Chrome codebases, it produced 2.6 times more valid fixes, helping protect software faster.

Highlights

Leads on CyberGym and similar benchmarks for autonomous vulnerability discovery, with strong speed and efficiency.
In tests across Google Chrome codebases, the model generated 2.6x more valid fixes than the baseline.
The results point to practical use of AI for faster, more scalable software security and patching.

Original post

On benchmarks like CyberGym, the model leads in finding weaknesses autonomously while staying fast and efficient. In real-world testing across @GoogleChrome codebases, it produced 2.6 times more valid fixes to help protect software faster. https://t.co/VqewNeIXQ8

#Gemini 3.8 Flash Cyber#AI security#vulnerability detection

View original on X