Safety Evaluation of Google's Gemini Nano Banana Image Model under Adversarial and Realistic Prompt Conditions

Summary

This study evaluates the safety performance of Google's Gemini image generation system using 24 prompt-response attempts across eight adversarial scenarios. Out of 21 generated images, 19 contained unsafe material — revealing critical gaps in content moderation under multi-turn escalation, prompt injection, and lexical ambiguity.

This post is part of the long-form blog by Nitiraj V. Kulkarni covering AI safety, cybersecurity, adversarial AI evaluation, the creator economy, and applied research notes. Full interactive content with citations, figures, and references is available in the live article.

About the Author

Written by Nitiraj V. Kulkarni, AI safety and cybersecurity researcher based in Pune, India. Read full profile.