Planning for Secure by Design AI is one of six short AI courses I bought from ISC2 in May 2026, and at $64 it is the bundle’s pass through the AI development lifecycle from the security side. It starts with traditional cybersecurity controls applied to AI use, moves through the legal and regulatory environment, spends real time on the data risks and dataset-protection problems that make AI a different shape of security target than a traditional system, and closes briefly on ethics and on AI applied to traditional cybersecurity work. Nothing in it is particularly surprising, and nothing in it is outright wrong.

What It Covers

The lifecycle frame is the structure. The cybersecurity section is the conventional one: securing the systems on which AI runs, the people who run it, and the processes around it – a straightforward extension of standard security thinking to a new workload. The regulatory section is a much lighter pass than Aligning with Global AI Regulations ; take Aligning if you want the wider tour. The data section is where the course has the most to say, and rightly so: training an AI requires vast datasets, the assembly of those datasets routinely sweeps in personal and health information that is not supposed to be there, and the dataset itself is then a high-value target that needs at least as much protection as the model, often even more. Ethics and AI-in-defensive-cybersecurity are brief, conceptual closes.

The substantive piece worth a separate look is the treatment of de-biasing. The course frames it as a question of how to address biases that result from unfiltered human input, and suggests it may be necessary to adjust the training data to reflect broader diversity. That recommendation can be taken two ways. The first reading is the human-thumb-on-the-scale one: put your judgment over what the AI would otherwise produce from the data it was given. That reading strikes me as suspect; you have stopped training a model on reality and started training it on what you wanted reality to look like, which is a different exercise with different downstream problems.

The second reading is the legitimate one. There are real cases where a particular demographic group is underrepresented in the input data, and the model’s predictions about that group are correspondingly less accurate – not by malice, just by sample size. Fixing that is a real problem, and the answer is genuinely subtle. It should not be confused with the well-publicized incidents of AI trained on unfiltered web input promptly declaring itself a controversial mechanized historical figure .

That incident actually shows both failure modes at once – the unfiltered training corpus that put the raw material into the model in the first place, and a later system-prompt change that told it not to shy away from politically incorrect claims. Dataset curation and instruction failed in concert. Both, at root, are the same ethical question. Do we train an AI on the sum total of human knowledge, accepting everything that comes along for the ride, or do we deliberately curate so the result is “better” than the average human? Asked plainly, everyone will answer that they want the AI to be better than the average human – particularly anyone who has spent any time around the average human, or worse, the below-average ones. What “better” means is the much harder question, and not one a $64 course is going to resolve.

What you get out of a model bears a very strong relationship to what you put into it. At small scale that is obvious. At training-corpus scale it is much harder to see.

Verdict

Three stars. Nothing here was new to me, nothing was wrong, and the lifecycle framing is useful as a checklist of what to think about before starting an AI project. I used an example from the de-biasing section, but the course is full of subtle dilemmas like it. AI technologies are magic, nor are they (yet) fully intelligent. They present a useful approximation of intelligence because their training data has been carefully curated to do that. Lacking the careful curation on known datasets, AI reasoning is just as susceptible to illogic, bias, and mistakes as the humans who wrote the training data in the first place.

A serviceable conceptual overview. Do not expect implementation guidance beyond the broad outlines, but the course will tell you what to watch out for before you trip on it yourself.

Part of The ISC2 AI Course Bundle . More from the bundle: AI for Cybersecurity · The Evolving Cybersecurity Workforce · Foundations of AI · Aligning with Global AI Regulations · AI Security: Managing Overconfidence