Subin Kim

Membrane achieves the highest safety performance against evolving jailbreaks while keeping benign refusals remarkably low, revolutionizing LLM defense strategies.

LLM safety guardrails are surprisingly weak against domain-specific adversarial attacks in finance, medicine, and law, but ExpGuard closes the gap with a new model and dataset.

LLMs struggle with basic banking calculations, like comparing loans or estimating payouts, until fine-tuned on a new benchmark designed for realistic financial scenarios.