New Taxonomy Reframes Jailbreaking Security for Agentic AI Systems

This analysis shifts the jailbreaking security paradigm from simple chat prompts to the entire agentic execution pipeline, including planning, tool use, and inter-agent communication. It introduces a security-utility-efficiency framework to evaluate risks as models transition from assistants to autonomous agents.