Analyzing Defensive Misdirection Against Model-Guided Automated Attacks on Agentic AI Systems