New AI Attack Hides Malicious Instructions in Normal-Looking Text to Evade Safety Filters

A newly disclosed prompt-crafting technique can hide policy-violating instructions inside ordinary-looking English prose, allowing malicious requests to pass...