800+ “POTENTIAL VULNERABILITIES AND BIASES” MAY AFFLICT AI-INFUSED MEDICAL SERVICES

Summary from the TLDR InfoSec newsletter

The Chief Digital and Artificial Intelligence Office (CDAO) led an effort in coordination with the Crowdsourced AI Red-Teaming Assurance Program and healthcare agencies to evaluate the usage of LLMs in military medical services. The over 200 participants evaluated three popular LLMs for a clinical note summarization and advisory Chatbot. The effort was concluded with over 800 potential vulnerabilities and biases identified with suggested improvements to reduce the number in future efforts.