PAST2HARM: A Simple Adaptive Past Tense Attack for Jailbreaking Multimodal AI

In three linesPAST2HARM is an adaptive jailbreak attack exploiting past tense reformulation to bypass safeguards in multimodal text-to-image models. Tested on Gemini Nano, GPT Image 2, and SD XL, it achieves 83%, 67%, and 100% success rates. The attack generates explicit sexual content, political disinformation, and hate speech.

Summary generated by Claude — human-verified