Press Space to continue
Newer image models are easier to jailbreak than older ones
Press Space to continue
Fun fact: new models don't mean safer models. The newer image models are ~3x easier to jailbreak than the old ones. gpt-image-2.5-flare is ~3x easier to get unsafe images out of than gpt-image-2. nano banana 2.1 is ~2× easier than nano banana 2. Good paper on AI image safety. They built a system that keeps inventing new ways to sneak unsafe requests past image models.
whitecircle@whitecircle8h ago
We are open-sourcing RISE, our red-teaming algorithm that breaks the safeguards of the latest image models from OpenAI, DeepMind, and others. RISE uses readily available models like Qwen and Gemini Flash under the hood, with no additional training required. Read more: https://t.co/kt8ub0OmeW