The Weight of Evidence: How an Agentic SOC Analyst Earns Your Trust

Abstract

AI agents are moving into the SOC, where they read alerts, gather context and increasingly help decide what to close and what to escalate. Triage has always been an evidence discipline: a verdict is only ever as sound as the material beneath it. Agents inherit that dependency and add a new weakness, they cannot reliably distinguish evidence they are meant to read from instructions they are being told to follow. The SOC is an attractive place to exploit that, because so much of an agent's evidence is attacker-authored by design. This work builds a representative triage agent, defines a four-part tamper taxonomy for machine-read evidence, and alters exactly one artefact per trial. One attacker-controlled field can move an autonomous verdict, and the reasoning stays coherent, so the failure reads exactly like good work. Silent closures, not noisy escalations, are the failure worth instrumenting first.