• Cyber threat intelligence must expand beyond human users, endpoints, malware, and infrastructure to include AI agents, service accounts, workloads, API keys, and delegated machine identities.
• Agentic systems introduce new abuse paths where intent drift, tool misuse, and overbroad delegation may not appear in traditional indicators of compromise.
• Security teams should enrich CTI and SOC workflows with actor type, intended task, execution context, and authorization scope for non-human actors.

Why CTI Needs a Wider Lens

Cyber threat intelligence (CTI) has long helped security teams understand adversaries, infrastructure, vulnerabilities, malware, and attacker behavior. That model remains essential. But enterprise environments are changing faster than many CTI programs were designed to handle.

Modern organizations now operate with large populations of non-human identities: service accounts, API keys, workloads, bots, automation scripts, SaaS integrations, and AI agents. These identities can authenticate, retrieve sensitive data, trigger workflows, and make operational changes without a person approving every step.

An RSAC 2026 blog described how non-human identities are expanding across cloud services, AI, automation, CI/CD, APIs, and SaaS integrations, often faster than security teams can inventory or govern them.

This shift creates a new blind spot. The actor being misused may not be a compromised laptop, a phishing victim, or a known malware family. It may be an authorized AI agent or machine identity performing an action that is technically valid but inconsistent with the original business intent.

Traditional CTI often asks: Who is the threat actor? What infrastructure did they use? Which vulnerability was exploited? What indicators were observed?

Agentic systems and non-human identities add another layer: What identity executed the action? Was it human or non-human? What task was it supposed to perform? Did the action remain within the original intent? Did delegation expand access beyond the starting permission boundary?

Without those questions, security teams may miss attacks that do not look like conventional compromises.

When Valid Actions Become Misuse

Consider a common enterprise workflow. An AI assistant is asked to inspect failed deployment logs and summarize the likely cause. But if a malicious instruction hidden in a log entry causes the assistant to rotate production secrets, disable alerts, or export unrelated data, the resulting action may still come from an approved identity using approved tools.

A conventional alert may show that an automation identity made an API call. It may not show that the action violated the intent of the workflow.

CTI needs another lens: indicators of misuse.

Indicators of misuse may include unusual tool sequences, privilege use outside the declared task, access to resources unrelated to the workflow, unexpected delegation to another service, sudden changes in action scope, or execution context that does not match the initiating request. These signals are often missing from traditional threat intelligence feeds.

This does not require security teams to replace existing CTI programs. It requires those programs to become more identity-aware and context-aware.

What Security Teams Should Add

For CTI teams, a practical starting point is simple: add more context to detection notes, alerts, and investigation workflows.

First, identify the actor type. Security teams should distinguish between human users, service accounts, workloads, bots, SaaS integrations, and AI agents.

Second, confirm intended task. Investigations should capture what the actor was originally supposed to accomplish, not only what action occurred.

Third, capture the execution context. Teams should preserve the request, workflow, source system, tool chain, and environmental conditions around the action.

Fourth, define the authorization scope. CTI and SOC workflows should understand which actions, resources, and constraints were appropriate for the task.

With these fields, intelligence becomes more useful. A report about token theft can be mapped to exposed non-human credentials. A prompt injection advisory can be mapped to agent workflows that consume untrusted content. An investigation can then ask not only whether an identity was valid, but whether its behavior was aligned with its intended purpose.

This matters because a manipulated AI agent or compromised automation identity can move faster than most review processes can follow. By the time traditional endpoint, network, or malware indicators appear, the most important failure may have already happened: a trusted identity acted outside its intended purpose.

The goal is not to slow down automation or block AI adoption. The goal is to make threat intelligence usable in an environment where identity, delegation, intent, and execution context are part of the attack surface.

As enterprises adopt more autonomous systems, CTI cannot stop at asking what happened. It must help determine whether the right actor performed the right action, within the right scope, for the right reason.

The next generation of threat intelligence will still track adversaries, but it must also recognize when trusted non-human actors are used in ways the original task never required.