Agentic triage

A triage model where an AI system can gather evidence, call tools, and decide what to inspect next during an investigation. It goes beyond summarisation or scoring because the system participates in the investigation loop and adapts its actions as new context appears.

Expanded Definition

Agentic triage describes a security investigation pattern in which an AI agent does more than rank alerts. It can query systems, inspect logs, request additional evidence, and adapt the next step of the investigation as new context appears. That makes it materially different from static scoring, summarisation, or rule-based routing, because the system participates in the investigative loop rather than only annotating it. In practice, the term sits at the intersection of AI security, SOC operations, and workflow orchestration, where tool access and decision authority must be tightly bounded. Guidance in the OWASP Agentic AI Top 10 and the NIST AI Risk Management Framework helps frame the risks that emerge when an AI system is allowed to decide its own next action.

Definitions vary across vendors on how much autonomy still counts as “triage” versus “automation,” and no single standard governs this yet. NHI Management Group treats the term as valid only when the system can both act on evidence and change its investigative path based on what it discovers. The most common misapplication is calling a simple alert classifier “agentic triage,” which occurs when a model only assigns priority scores but cannot gather evidence or initiate the next investigative step.

Examples and Use Cases

Implementing agentic triage rigorously often introduces governance and containment overhead, requiring organisations to weigh faster evidence collection against tighter tool restrictions, auditability, and human oversight. The practical challenge is not whether the agent can inspect more quickly, but whether its actions remain explainable and safe enough for operational use.

• A SOC assistant receives a suspicious sign-in alert, checks identity telemetry, pulls recent endpoint events, and then decides whether to escalate to a human analyst.
• An AI workflow inspects a possible NHI compromise by querying token issuance logs, API gateway traces, and secret-rotation history before recommending containment.
• A phishing investigation agent opens related messages, searches for identical indicators across mailboxes, and chooses the next evidence source based on whether the campaign appears targeted or broad.
• A cloud security assistant reviews a high-risk configuration finding, gathers IAM context, and verifies whether the exposure is actually exploitable before creating a ticket.
• Threat-hunting teams use agentic triage to move from first alert to enriched case notes, but only after placing guardrails around tool use and prompt injection exposure, as reflected in the MITRE ATLAS adversarial AI threat matrix.

Why It Matters for Security Teams

Agentic triage matters because it changes where trust is placed. When the system can choose what to inspect next, errors are no longer limited to bad rankings; they can become bad investigative paths, missed evidence, or unsafe tool actions. That creates new requirements for authorization boundaries, logging, replayability, and review of every action the agent takes. For teams governing AI-enabled security workflows, the key question is whether the agent is merely assisting an analyst or is effectively making operational choices that influence response outcomes. The NIST SP 800-53 Rev 5 Security and Privacy Controls is relevant here because the surrounding controls must cover access, audit, and system integrity, while agent-specific threat models such as CSA MAESTRO agentic AI threat modeling framework help teams reason about unsafe autonomy.

In identity and NHI contexts, the term becomes especially important because an agent that can inspect tokens, service accounts, and access trails is operating inside the identity plane as much as the security plane. Organisations typically encounter the operational consequences only after an agent has taken an unhelpful investigative branch or queried the wrong system, at which point agentic triage becomes operationally unavoidable to correct.

Related resources from NHI Mgmt Group

• Agentic Supply Chain
• What is Agentic AI and how does it differ from traditional generative AI?
• What NHI types do Agentic AI systems typically use?
• Why does Agentic AI dramatically increase identity sprawl?