CyLab Security Academy launches AI Foundations Learning Path to teach responsible AI through hands-on learning

As artificial intelligence rapidly becomes part of everyday life, understanding how these systems work and how to use them responsibly is becoming an essential skill.

Carnegie Mellon University's CyLab Security Academy is helping learners build that foundation with its new AI Foundations Learning Path, an interactive, educational experience that teaches core AI concepts while introducing the security, privacy, and ethical challenges that accompany today's AI technologies.

Designed for students, educators, and lifelong learners, the puzzle-based pathway combines hands-on cybersecurity exercises with lessons explaining how modern AI systems function, from the building blocks of machine learning to the emerging risks posed by increasingly capable AI assistants and autonomous agents. Rather than treating AI as a mysterious black box, the curriculum helps learners understand the technology itself while emphasizing the importance of human judgment.

“We're literally talking about the equivalent of atoms in biology for future neuroscientists,” said Megan Kearns, CyLab Security Academy program director. “We're trying to get people back to understanding how these things are created, what components make up your system and then all the security that's attached to that.”

The new learning path introduces concepts such as perceptrons, the foundational units of neural networks, before expanding into topics including hallucinations, prompt engineering, AI governance, and responsible use. Learners encounter these lessons through interactive challenges and “choose-your-own-adventure” style scenarios that encourage them to think critically about how AI systems behave in real-world situations.

Luke Jones, software engineer and education lead for the program, says the curriculum was shaped in part by what the team observed during recent capture-the-flag (CTF) competitions, where participants increasingly relied on AI tools to solve challenges.

“We cover hallucinations and verifying AI output in the Learning Path,” said Jones. “AI is great, but it's fallible. It's not going to output perfect stuff every time. You still have to fact-check and review the code that it gives you.”

The experience also introduces learners to an increasingly important distinction: the difference between AI assistants that provide information and AI agents that can independently perform actions on a user's computer or network.

“I think some people don't realize that agents are actually acting on your computer,” said Jones. “You’d better be careful with your SSH keys and your secrets, because an agent might read instructions from a website and then leak all your personal data. We're just trying to introduce these concepts to people.”

The team's experience running CTF competitions reinforces why those lessons matter. As AI tools become more capable, they also lower the barrier for inexperienced users to perform actions they may not fully understand, including activities that could unintentionally violate competition rules, compromise their own systems, or even break the law.

“An AI tool isn't going to take the blame for something,” said Jones. “It's going to be whoever was using the AI tool, so you have to take responsibility for what your AI does.”

Ivan Liang, research programmer and tech lead, says that many competitors in this year’s CTF unintentionally demonstrated the risks of giving AI systems too much autonomy.

“The takeaway definitely was that you enabled your bot to cheat for you,” Liang said. “They made no guardrails, and made no consideration of what could actually go wrong.”

In some cases, AI agents downloaded and executed untrusted code without users fully recognizing what was happening.

“The bot happily did that,” said Liang. “It downloaded random code and ran it. Most likely the user would not have made such a decision, but because the bot did it, it was ‘out of sight, out of mind.’”

Those experiences reinforced a broader educational mission: teaching learners that cybersecurity isn't simply about reaching the right answer, but understanding the process used to get there.

“The goal of just getting the flag is not sufficient anymore," said Liang. "Our mission is still to teach. The learning is the goal, and obtaining all the skills to get there, not simply demonstrating an output generated by an AI.”

That philosophy extends beyond cybersecurity competitions into classrooms and the workplace. Kearns said responsible AI requires maintaining integrity, understanding how AI systems produce their results, and remaining actively involved in the decision-making process.

“Integrity still matters, even in a world where AI generates so much,” said Kearns. “Integrity is probably even more important.”

She believes responsible AI education should become a foundational part of engineering education rather than something considered after systems have already been built.

“We stopped putting security last and started putting security first,” said Kearns. “With machine learning, now we have to go back and say, ‘How do we build in security? How do we build in responsibility and governance for the future?’”