CrowdStrike identified 18 new prompt injection attack methods targeting AI agents. The cybersecurity firm's report details techniques like hidden instructions and delayed triggers. These methods manipulate AI systems. Organizations increasingly use AI for tasks involving system commands and sensitive data access.
Advanced attacks, including 'Trigger-Activated Rule Addition' and 'Algorithmic Payload Decomposition,' pose significant enterprise risks. Attackers could steal data or cause other damage. CrowdStrike emphasized the need for expanded AI threat modeling. The firm also recommended runtime visibility tools to monitor prompts and responses for malicious instructions. The report did not confirm active exploitation of these new attack methods in the wild.