New AI Audit Method Stops Harmful Content | Tech News

A groundbreaking new method lets researchers spot AI models secretly tweaked to generate illegal content—like child abuse material—without ever seeing the harmful output. Traditional testing fails here because creating such content is illegal. Scientists at MIT and Thorn developed a technique that analyzes internal code changes, specifically low-rank adaptations, to detect malicious modifications with 100% accuracy. This breakthrough empowers platforms to block dangerous models before upload, closing a major loophole and offering a vital safety net for AI in the open-source era. Listen in comfort:Get a discount on a Soli Pillow: http://solipillow.com/discount/dnn. Advertise on DNN:[email protected] This is an automated, high-level news summary based on public reporting.Report issues to [email protected]. View sources & latest updates:https://sources.thednn.ai/a6308f1570ad2502