Simple image edits still bypass commercial AI moderation
Researchers tested three commercial moderation services with seven inexpensive image transformations. Colour inversion, grayscale and related edits caused harmful images to be accepted while remaining recognizable to people, with especially weak results for multimodal and self-harm content.
Why it matters: A multimodal foundation model does not become a dependable security boundary merely because it understands more context than a traditional image classifier.