Research Detects Prompt-Injection Hijacks via Model Probabilities
Monitor using output probabilities detects some hijacks on AG News classifier but misses most when attackers add certainty claims.
Stories per day · 30 days
1 stories, newest first · one card per story