Researchers Map AI Threat Landscape, Risks

This article discusses the systemic security risks of large language models (LLMs), focusing on the "black box" nature of LLMs as identified in a BIML report and complementary NIST taxonomy work. It highlights 81 risks (23 tied to black-box issues), the difficulty of defending generative models against adversarial prompts, and calls for regulatory accountability for foundation-model providers while noting current mitigations are limited and a cat-and-mouse dynamic persists.