Jindong Gu

Autoregressive image generation can be made significantly safer by iteratively refining codebooks based on the model's own judgment of harmful outputs.

SeClaw reveals that existing benchmarks fall short in capturing the complexities of agent behavior, enabling a more nuanced evaluation of security risks in autonomous systems.

Adversarially finetuning CLIP using a pretraining-inspired recipe with web data and feature regularization yields significantly better zero-shot robustness across diverse datasets than standard adversarial training.

Coding agents are vulnerable to a new class of stealthy, automated prompt injection attacks via poisoned skills, achieving high success rates even in realistic software engineering tasks.