Break One Model, Break Them All

Adversarial Attacks Transfer Across Architecturally Different LLMs and No One Understands Why

digital-llm-adversarial-transferability