A new jailbreak called CollageAttack hides harmful prompts in spatially arranged text fragments within a single input, bypassing safety filters in text-to-image models and succeeding up to 86% of the time. It shows that…

#T2ISafety #AIsecurity #Jailbreak
https://arxiv.org/abs/2609.38253

New work proposes ICER, a black-box framework for red-teaming text-to-image models, using an LLM rewriter for fluent adversarial prompts and in-context experience replay to reuse successful patterns.

#T2ISafety #RedTeaming #PromptEngineering #AIRisk
https://arxiv.org/abs/2411.16769