Mastodon Feed
Boosted by cstanhope@social.coop ("Your friendly 'net denizen"):
bradlarsen@infosec.exchange ("Brad Larsen") wrote:
This is a really neat attack: craft an image that, when downscaled by a multimodal AI system, contains embedded text not visible to a human in the original image, which is used for prompt injection
https://blog.trailofbits.com/2025/08/21/weaponizing-image-scaling-against-production-ai-systems/