Geometric analysis proves global 'unsafe direction' defenses for text-to-image models are fundamentally flawed

T2I safety filters using toxic subspace methods are formally broken; deploy replacements before adversarial exploits go mainstream.

→ arXiv cs.AI