Safety considerations call for deployment of autonomous ground vehicles in defense and high risk zones for transport of goods from one point to another. Such vehicles face the threat of an intelligent autonomous adversary that may disrupt the transfer of material. This article investigates the challenges involved in autonomous protection of a delivery agent, via a land-based rescue agent, before interception of the delivery agent by the adversary occurs. In particular, we study how effectively an adversary equipped with a vision sensor can be handled by an autonomous rescue agent operating without vision support and relying only on wireless communication with the delivery agent. Taking capabilities and weights of the three vehicles into account, the delivery agent is assumed to be the slowest while the adversary operates at the highest speed among the three vehicles. A geometric framework based on Apollonius circles is proposed to analyze the interaction between the delivery and rescue agents. The adversary's speed and its moves (based on the direction of the delivery agent) are taken into account, along with the Apollonius circles for the rescue-delivery agent pair, to determine the possibility of capture. Regions in the plane where the delivery and rescue agents can meet, prior to a capture by the adversary, are obtained to compute safe regions for the delivery agent. Algorithms adopted by the delivery agent, rescue agent, and the adversary are described. We, then, explore the challenges in rescue of multiple delivery agents from a vision-guided adversary by introducing additional rescue agents. In particular, we study protection of k delivery agents (from an adversary) via k rescue agents. Algorithms to compute 1) multiple meeting points, one each for a delivery agent-rescue agent pair and 2) the strategy of the adversary to capture any one of the k delivery agents are presented. Experiments with multiple agents show that the delivery and rescue agents can execute their strategies using simply low-end microcontrollers without external memory. © 2005-2012 IEEE.