AI RESEARCH
SALLIE: Safeguarding Against Latent Language & Image Exploits
arXiv CS.AI
•
ArXi:2604.06247v1 Announce Type: cross Large Language Models (LLMs) and Vision-Language Models (VLMs) remain highly vulnerable to textual and visual jailbreaks, as well as prompt injections (arXi:2307.15043, Greshake, 2023, arXi:2306.13213). Existing defenses often degrade performance through complex input transformations or treat multimodal threats as isolated problems (arXi:2309.00614, arXi:2310.03684, Zhang, 2025