ai safety benchmarks

Google's Gemini 2.5 Flash AI model shows a **4.1% decline in text-to-text safety** and a **9.6% drop in image-to-text safety**…