Microsoft researchers discovered that poisoned AI models exhibit normal behavior until specific trigger words cause them to “blow up” or dramatically change responses. PCWorld reports that unlike ...