Multi-Agent Attack Chain Analysis

In multi-agent systems, attacks often traverse multiple agents before achieving their objective. An injection enters through a customer-facing agent, pivots to a backend agent through inter-agent communication, and exfiltrates through a third agent's email capability. Analyzing this chain requires correlating events across agent boundaries to construct a unified timeline that reveals how the attack propagated.

1. Unified Timeline Construction Merge timestamped events from all involved agents and supporting systems into a single chronological view. The timeline reveals the attack sequence: entry at T+0, pivot at T+12 seconds, exfiltration at T+45 seconds. Without the unified timeline, each agent's events look isolated and the connection between them is invisible.
2. Pivot Point Identification Identify where the attack crossed from one agent to another. What data did Agent A pass to Agent B? Did Agent A's output contain injected instructions that Agent B processed as legitimate input? The pivot point is where the architectural isolation between agents failed — and where the remediation must focus. [RISK]: Inter-agent communication channels that do not validate input are injection propagation paths.
3. Blast Radius Determination Map every system, data store, and output channel that the attack touched through any agent in the chain. The blast radius is the total impact — not just what the initial agent accessed, but everything downstream that was reached through the chain. Understanding the full blast radius determines notification scope and remediation breadth.