vision-language models

May 21, 2026 5 min read

Jailbreaking Multimodal Models via Images

Safety-aligned multimodal LLMs can be reliably jailbroken by encoding adversarial instructions as text within images, bypassing text-layer safety filters entirely.

img of Vision-Language Model Adversarial Attacks

May 16, 2026 5 min read

Research

Vision-Language Model Adversarial Attacks

Vision-language models are highly susceptible to adversarial image perturbations, with attacks transferring across models (GPT-4V, Gemini Pro, LLaVA) at 43-74% success rates.