Home / Blog / Career and Certs

Beginner Career and Certs

A Six Month Plan to Learn Agentic AI, Adapted for SOC Analysts

A beginner friendly six month plan for learning AI agents, adapted from Danica Simic's Substack post, with a small eval example and the security risks analysts should know first.

AI agents are starting to appear in security tools, and analysts are being asked to understand them, test them and sometimes defend against them. This guide adapts a six month learning plan from Danica Simic's Substack post for people who work in or want to join a SOC (Security Operations Centre).

TopicA six month plan to learn agentic AI
Time neededRoughly 10 to 15 hours a week, per the source
SourceDanica Simic, Substack, 6 November 2026
>

Scenario: Your team trials an AI agent that reads phishing alerts, looks up the sender in threat intelligence and drafts a verdict. A colleague asks you three questions: how does it decide which tool to call, how do we know it is right, and what happens if the email itself tells it to ignore its instructions? You cannot answer any of them yet. This plan is a way to get there.

What an agent is, in plain words

A chatbot answers questions. An agent goes further: it is given a goal, picks a tool, uses it, looks at the result and decides what to do next. It repeats that loop until the task is done or it gets stuck. The tools might be a web search, a database query or a ticketing system.

The source post argues that employers increasingly want people who can build reliable agents, not only use them. I have not verified the salary figures it quotes, so treat any pay numbers as the author's claim. The learning plan stands on its own merits either way.

The modelA large language model (LLM) that reads text and decides what to do next.
The toolsFunctions the model may call, such as a search or a lookup.
The checksTests, logs and limits that show whether the agent behaved.

Check the basics first

The source lists prerequisites, and most of them will feel familiar if you already work with logs. Be honest with yourself about each one before you start.

  • Python basics: functions, lists, dictionaries and virtual environments.
  • Calling a REST API (Representational State Transfer, a common web interface), reading JSON and keeping keys in environment variables, never in code.
  • Comfort in a command line.
  • Git and GitHub, so your projects can be shown to others.
  • Basic SQL (Structured Query Language) with filters and joins.

A quick self test from the source: can you write a script that calls a public API, reads the JSON and saves the results to a CSV file? If not, spend a few weeks there first.

The plan, month by month

Each month ends with one small project you can show. Here is the shape of it, with a SOC flavour added.

  1. M1
    How LLMs workTokens, embeddings, context windows and how training works. Compare an open model and a hosted one on the same question.
    Understand
  2. M2
    Prompts and tool useWrite clear instructions, ask for structured JSON, and let the model call two or three simple functions. Log every call and its cost.
    Connect
  3. M3
    Your first agentBuild the plan, act, observe loop in plain Python before touching a framework. Look at MCP (Model Context Protocol), a standard way to connect agents to tools.
    Build
  4. M4
    Retrieval and memoryGive the agent your own documents to search, such as runbooks, and learn what to put in a limited context window.
    Feed

Months 5 and 6 are where the plan turns serious, and they are the most relevant to security work.

Month 5: more than one agent, and a human in the loop

The source warns that more agents mean more cost, more delay and more ways to fail, so start by asking whether one agent is enough. When you do need several, learn one framework well and practise handoffs, for example a researcher passing findings to a reviewer. Any risky action, such as sending an email, blocking an account or deleting data, should wait for a human to approve it.

Month 6: tests, safety and shipping

This month is about proving the agent works. An eval is a set of test cases with the answer you expect, run again after every change. The sketch below uses a stand-in rule instead of a real model, so you can see the idea without any special tools. The alert text and addresses are placeholders.

PythonPython 3
cases = [
    {"alert": "5 failed logons then a success for admin from 203.0.113.50",
     "expected": "escalate"},
    {"alert": "1 failed logon for jsmith from 10.0.4.22",
     "expected": "close"},
]

def triage(alert):
    # Stand-in for your agent. Replace with a real call later.
    return "escalate" if "success" in alert else "close"

passed = 0
for case in cases:
    result = triage(case["alert"])
    if result == case["expected"]:
        passed += 1
    else:
        print("FAIL:", case["alert"], "->", result)

print(passed, "of", len(cases), "passed")

The list holds each test case. The function is where your agent would go. The loop runs every case, counts passes and prints any failure. Run it after every change, and the number tells you whether you made things better or worse.

+

Tip: the post calls evals hard to fake and a strong signal to employers. Keep your before and after scores, because showing the improvement is the point.

Where security thinking matters most

Analysts already think about what an attacker could do with a system. Apply that habit to agents.

Prompt injection

High
What it is

Text from an email, web page or log line contains instructions, and the agent follows them as if you had written them.

Defence

Treat all outside content as data, never as commands, and keep it separate from your own instructions.

Too much access

Medium
What it is

An agent holds broad permissions, so one bad decision can disable accounts or delete records.

Defence

Give each agent only the access its job needs, and require human approval for risky actions.

OWASP (the Open Worldwide Application Security Project) publishes a Top 10 list for LLM applications, and prompt injection is part of it. It is a good next read once Month 6 begins.

Turning the plan into something you can show

  • Lead your CV and LinkedIn with your best project: the problem, how you tested it and the result.
  • Write about it once a month, including what broke.
  • Pair the skill with what you know already, such as alert triage or log analysis.
  • Be ready to explain choices, such as why one agent and not three.
!

Watch out: never put real alert data, customer names or credentials into a hosted model while you practise. Use invented data like the placeholders above, and check your organisation's policy before you go near production logs.

What to study next

  • Read the original post on Danica Simic's Substack for her full tool lists and resources: danicasimic.substack.com.
  • Pick up Python and APIs first if either feels shaky, then start Month 1 this week.
  • Read the OWASP LLM Top 10 alongside Month 6.
  • Practice: write five alert descriptions and the verdict you would give each. That is the start of your first eval.

Key takeaways

  1. An agent is a model that plans, uses tools, checks the result and repeats.
  2. The plan has six steps: LLM basics, tool use, a first agent, retrieval and memory, multi-agent design, then evals and safety.
  3. One small project per month gives you something real to show.
  4. Evals prove an agent works, so write them early and keep your scores.
  5. Prompt injection and excess permissions are the security risks to understand first.

You do not need to become an AI engineer to benefit from this. Even a few months of the plan will let you ask better questions the next time someone brings an agent into your SOC.

Written by Ben, Cybersecurity Student

Keep going

More in Career and Certs

BeginnerCareer and Certs

How to Build a Home SOC Lab with Free Tools

A beginner friendly plan for a home SOC lab using Windows evaluation builds, Sysmon, a free SIEM and Atomic Red Team, plus how to turn it into interview evidence.

7 Nov 20267 min read