Initialization Anchoring Weakness in Feedback-Based Agent Planning

Feedback-based agents exhibit an initialization anchoring weakness where the first feedback round corrects 46% of adversarial directions, but effectiveness drops to 7% by the third round. The proposed InitAnchor framework exploits this via attacker-controlled materials that leverage contextually plausible shifts and persistent trajectory directions.