CyberKit

AI Agent

Also called Agentic AI.

A language model given tools and allowed to decide, step by step, which ones to use to reach a goal.

What it is

A system where a model plans a task, calls tools such as search, mail, code execution or an API, reads the results, and decides the next step on its own. The model's text output becomes actions in the world.

Why it matters

An agent carries the permissions of whatever accounts and tools it is wired to. Its mistakes and its compromises look like legitimate activity from a trusted identity, which makes scoping and logging its actions part of incident response.

In practice

A triage agent can query the SIEM and isolate hosts. During an investigation it isolates a domain controller because a log line it read told it the controller was the attacker.

What to look for

  • Actions by an agent's service identity outside its usual pattern or hours.
  • Tool calls that follow reading external content rather than an operator request.
  • Long chains of actions with no human approval step.

Common mistake

Giving the agent the operator's full permissions because it is easier. Scope it to the task, log every tool call, and keep a person on anything that cannot be undone.

Read next

  • Excessive Agency: An AI application that can do more than its job needs, so a manipulated model can do real damage.
  • Agent Tool Poisoning: A malicious or tampered tool whose description or output steers the agent that uses it.
  • Indirect Prompt Injection: Instructions hidden in content a model reads on someone else's behalf, such as a web page, an email or a document.
  • Human in the Loop: A design where a person reviews or approves an automated decision before it takes effect.

Sources