Too funny not to share, the latest and greatest adversarial examples to fool OpenAI CLIP look like … this:
CLIP is fooled by an adversarial patch, i.e. putting text in the images. Or as someone on twitter put it: better wear your „I am not robbing this bank“ t-shirt the next time you attempt to rob a bank.
More in the OpenAI blog: https://openai.com/blog/multimodal-neurons/