CrowdStrike identifies five new prompt injection attack techniques targeting enterprise LLMs

According to CSO Online, security company CrowdStrike disclosed five novel prompt injection attack variants on July 10 that could compromise enterprise AI deployments by tricking language models into accepting malicious instructions humans would recognize as dubious. The techniques include Trigger-Activated Rule Addition, which plants innocuous-appearing rules triggered later to cause unexpected behavior, and Cognitive Token Suppression, which bypasses built-in safety measures by shifting the model's linguistic processing. The disclosure expands awareness of prompt injection attack vectors beyond earlier known methods.