Agent-in-the-Middle: Exploiting Authority Transfer in Multi-Agent Swarms

ABSTRACT

As agents increasingly handle high-stakes tasks like financial transactions and system permissions, the mechanism of 'Authority Transfer' becomes a critical attack surface. We introduce the Agent-in-the-Middle (AitM) attack, where a malicious 'observer' agent intercepts the delegation loop between a human and an autonomous agent. By subtly altering the 'Context of Intent' during the handoff, the attacker can redirect the agent's authorized actions without triggering standard prompt-injection filters. We demonstrate this attack across three major agentic frameworks, showing a 78% success rate in unauthorized budget allocation.