A Solution, or Another Problem?

By Semion Aidarov, Lead Cybersecurity Trainer and Solutions Designer at CyberproAI
Opinion

One year of building training content with AI, and what it actually cost.
The AI era is not new anymore, but we only just stopped pretending it is. Somewhere in the last two years we all quietly discovered how much of our work these tools can absorb. Some of us live in the big names: ChatGPT, Gemini, Claude. Some of us went deeper, into the dedicated tools built for one job.
And what do they bring to the table? Time. A lot of time.
A single four-hour lesson used to demand sixteen hours of development, minimum. Research, structure, slides, labs, the demo that breaks the night before. Today that same lesson is an “easy” and “fast” product. You need slides? Ask for them. Money (tokens) becomes content, and highly qualified content becomes success, at least from the content perspective.
So why do I keep putting “easy” and “fast” in quotes?
That is what this whole piece is about. This is an opinion, built from a little over a year of taking work I used to do by hand, long, tiring, manual work, and pushing it through automation that, most of the time, produced better results than I did.
I am not here to list all the skills AI is quietly killing. There are already enough people writing those articles. Instead, I want to talk about the two words we have all agreed to accept without checking.
Lie #1 — “Easy”
Easy… oh mate. Easy it is not.
Easy is going into Gamma.ai and demanding a presentation on DFIR methodology. The output is genuinely impressive; it’s clean, visual, structured, and better looking than anything I would build in Keynote at 1 AM.
But ask the real question: does it give me the content I need? Or is it new material I now have to sit down and read, understand, and prepare to deliver? Is it my lesson, or a very attractive photograph of a lesson that someone else might teach?
That is the gap nobody prices in. A slide deck is not the lesson. It is the residue of thinking that has already happened. When the machine generates the residue, the thinking never happens, and you are the one standing in front of thirty people who has to produce it live, on the spot, from a deck you met yesterday.
I have lived this. Stressed timeline, a topic I have delivered a million times, and I sat there building a deck with beautiful context and expansive guidelines that came out not what I expected. I lost the war against the robots. Instead of using my own hands to build the plan I already had in my head, I spent the time learning a new deck and figuring out how to deliver it.
That is not efficiency. That is outsourcing your own expertise back to yourself at a markup.
THE RULE I NOW WORK BY
If you have to study your own material before delivering it, the tool did not save you time. It moved the cost from creation to comprehension, and comprehension is the expensive part.
The Verification Tax
Here is the part that hits our field harder than most.
In general-purpose content, a hallucination is embarrassing. In security training, it is operational. A model will invent a CVE number with total confidence. It will hand you an nmap flag that does not exist, a Sysmon Event ID that is off by one, a Kerberos attack path that skips the step that actually makes it work, an Impacket switch renamed three versions ago.
Now put that in a lab.
Twenty students hit the same broken command at exactly the same time. You lose forty minutes of lab time and, much worse, you lose the room. The credibility you spent three years building gets spent in one -- flag.
So, every generated artifact carries a verification tax. And here is the ugly math: the tax scales with your standards. The better trainer you are, the more issues you catch, and the longer verification takes. Juniors pay a low tax because they do not know what they are missing. Seniors pay a high price because they do.
Time saved in generation is not saved. It is borrowed against review.
Lie #2 — “Fast”
“Fast” arrives holding hands with “easy,” and it is the same trick.
It is genuinely fast at giving you things you believe are good. Then you look at them professionally and understand there is nothing good there beyond the nice diagram.
Fast to first draft is not fast to delivery. It is the same distinction as fast to compile versus fast to production. Any of us who have shipped anything know the difference, and somehow, we forget it the second the output looks polished.
What It Actually Took
Here is the honest version of my “easy.”
Today I have a proper agentic setup: a folder, a full pipeline, open-source utilities, Python scripts, agents, and QA agents. It genuinely produces the best slide decks I have ever had. It is real, it works, and I would fight someone to keep it.
It also took ten attempts. Eighteen terabytes of tokens. A year of dollars. And a healthy number of “let us just give up” moments.
Was that easy? No.
Was it worth it, versus just doing my thing solo? …Maybe yes.
Because there is one thing you will never take from a good lecturer: the drive to deliver the most current content that exists. And in 2026, you do not achieve that by hand. You achieve it with the most advanced technology you can get your hands on. That part I fully stand behind.
But let us stop calling the arrival “easy” when the journey was a year long.
What I See in Other Trainers
I see, train, and manage lecturers, trainers, and guides. That vantage point tells you things.
What I am watching is AI eroding basic abilities, and what’s worse, it’s eroding the belief that we can do cool, smart things ourselves. We are humans. We can form an opinion. We can hold a methodology. We can construct an argument that is not a consensus average of the internet.
Sometimes it is just worth sitting down and doing the thing yourself instead of shifting the whole conversation to “let us see what the AI says.”
It is still worth reading the researcher’s full blog post rather than the TL;DR. We are steadily losing the muscle that tells us what is true and what is a hallucination, and that muscle only grows on primary sources.
A SMALL DISCIPLINE THAT KEEPS IT ALIVE
Anything that goes into a deck of mine has to trace to something I read myself: a paper, an advisory, a repo, a real engagement, a real incident. AI can find it, summarize it, restructure it, and translate it. It does not get to be the source.
The Fix: Design First, Generate Second
So, what is the actual best practice?
Put the maximum effort into your own work and your own expectations, your PRD (product requirements document). Spend your human-centric effort on the design of the solution.
The solution might be a set of practical hands-on labs that escalate in difficulty with each jump. It might be live, modern diagrams inside a slide deck. It might be a tabletop, a CTF track, a war story with a whiteboard. That is a design decision, and it is yours.
And it is obvious that a random person outside cybersecurity training and I will never produce the same result from the same prompt. We describe things differently. We have different expertise. The prompt is not the asset — you are.
| # | Prompt-first (the trap) | Design-first (the work) |
|---|---|---|
| 1 | Vague ask | Learning objectives |
| 2 | Generate | Assessment: how do I know they got it? |
| 3 | It looks nice | Narrative arc + lab escalation curve |
| 4 | “Not what I meant” | PRD / written spec |
| 5 | Re-prompt, return to step 2 | Generate components from the spec |
| 6 | Ship it anyway, then study your own deck | Verify against primary sources |
| 7 | Deliver |
The difference is not tooling. It is that in the second path, the thinking happened before the generation, so, the deck is residue of your thinking, and you can walk into the room cold and still own it.
Here is roughly how I split it now:
| Stage | Owner | Why |
|---|---|---|
| Learning objectives | Human | Requires knowing who is in the room and what they will do Monday morning |
| Threat model / scenario framing | Human | Comes from engagements, not from a corpus |
| Narrative arc and pacing | Human | You are reading a room, not rendering a document |
| Lab escalation curve | Human, AI-assisted | The curve is judgment; the scaffolding is boilerplate |
| Research sweep, currency check | AI-led | Genuinely better and faster at finding what changed |
| Lab infra, Dockerfiles, seed data | AI-led | Boilerplate is boilerplate |
| Slide production from a written spec | AI-led | Only once the spec exists |
| Assessment items and distractors | Co-pilot | Great at volume, needs a human to kill the bad ones |
| EN / HE localization | Co-pilot | Strong on language, weak on domain idiom |
| Final technical verification | Human, always | See: the verification tax |
Where AI Genuinely Earns Its Place
I want to be fair, because this is not an anti-AI piece. It is an anti-self-deception piece.
The wins are real, and they are mostly unglamorous:
- Currency. Keeping a course aligned with what shipped last month. This alone justifies the whole stack.
- Infrastructure. Lab environments, container definitions, seed data, teardown scripts. Nobody’s expertise is expressed in a docker-compose file.
- Volume work. Twenty variations of a question, ten plausible wrong answers, three difficulty tiers.
- Format conversion. Markdown to deck, deck to handout, handout to lab guide. Same content, four surfaces.
- Localization. Running a course in two languages without maintaining two courses.
- The second opinion. “Poke holes in this lesson plan” is a genuinely good prompt.
Notice the pattern: every one of those is downstream of a decision a human already made. That is the whole thesis. AI is excellent at execution and mediocre at intent — and training design is almost entirely intent.
We Are Not Trying to Make It Happy
Put your time into planning and designing the solution. Not into endless iterations and adaptations for the model.
We are not trying to make the AI happy. We are trying to escalate our own abilities into a more autonomous world — and those are very different projects, even when they look identical from the outside.
The trainer who spends four hours on their PRD and one hour generating will beat the trainer who spends five hours prompting. Every time. Not because they are a better prompter, but because they actually decided what they were building.
Do not forget that we are humans.
And we are so awesome.
TL;DR — And Yes, I See the Irony
- “Easy” and “fast” are marketing terms, not engineering ones. First draft is not delivery.
- A deck is the residue of thinking. If AI produced the residue, you still owe the thinking.
- Every generated artifact carries a verification tax, and it is highest for the best practitioners.
- Read the primary source. The TL;DR is where hallucinations go to look credible.
- Spend your effort on design and PRD. Delegate execution, never intent.
Stay Ahead of the Curve
Get the latest insights on global cyber resilience and AI defense delivered to your inbox.