A new benchmark evaluation has shown that injection detection classifiers fail to flag most attacks when they are embedded in ordinary-looking data. The study tested 629 cases with 27 distinct attack patterns against multiple safety models.
log in to read full article