As we were creating our new AI security path, we spoke to leaders across healthcare, finance, automotive, manufacturing and insurance to understand what they were most worried about, and where we could help.
The results were remarkably consistent.
The AI risks security leaders worry about
1. AI is already in production across industries, departments and environments. What had been theoretically discussed at conferences in the recent past is now a reality that isn't entirely governed. "Everyone has AI systems in their production environments handling extremely confidential data," says Vlad Boldura, Senior Manager, Content Engineering at TryHackMe. "This is the reality of today."
2. And yet security teams have no coherent framework to protect any of it. As Vlad puts it: "They're asked to protect things they don't understand end-to-end. They just implement random small controls here and there."
The productivity gains from AI are real, but as Max Robertson, Senior Content Engineer at TryHackMe, notes: "All of these productivity gains are coming at a cost — and that cost is an exponential increase in attack surface."
3. Supply chain came up unprompted in almost every conversation. Leaders across industries said they don’t have sufficient control over their AI supply chain, and they're worried about it. The risk extends well beyond the security team: from developers, to DevOps, engineers and IT, most technical teams have not been adequately trained to think about what using AI daily means from a security perspective. "Every single engineer will need to understand this," says Vlad. "It's no longer reserved for the senior security researcher."
The training gap
Security leaders told us that the training available to address these concerns is either too expensive, too advanced to be inclusive across most teams, or too passive to build real skills. Nothing on the market covers the full threat landscape, at every level, with genuine hands-on AI interaction.
"You can watch a 45-minute lecture on prompt injection and understand the concept perfectly. Now sit down in front of an actual AI system with guardrails and try to extract the system prompt. Suddenly that lecture hits a wall,” Vlad told us.
One client put it plainly: their team could pass knowledge checks but couldn't reason to solve any real-world novel problem. That theory has nowhere to go, and no objective means of validation.
The other problem is how most training handles AI itself. "A lot of other training simulates AI because of the resources required to power it. But if it's simulated, it's deterministic, you always get the same output for the same input — and it's not really capturing the scope of what we're dealing with from a security perspective,” Max explains.
AI security is fundamentally non-deterministic. One user sends a prompt and gets one response. Another sends the exact same prompt and gets something completely different. Human language is the attack surface, and can be used in almost infinite permutations. To truly understand this, teams need firsthand experience, on a real system.
Managers own the AI security training gap
"A lot of managers are advocating for the use of AI because they want to see productivity go up,” Max underlines. “But if a security incident happens because a member of your team used AI technology that you were incentivizing, and it wasn't done in a secure way, there could be a compounding effect on brand equity and trust, not to mention legal and regulatory consequences, or impact on revenue.”
AI usage is now structural within teams, and requires organizational accountability to match it. Every engineer using an LLM to write code, every developer downloading models from repositories, every team integrating AI into their workflows is expanding the attack surface. But most of them aren't addressing these risks in the way they work.
"Most security leaders know they have an AI skills gap. They can feel it if they don't know it. They read the headlines, they see AI getting deployed across their org, and they know their team isn't ready." We’ve built a path to address this very gap: in security teams and the technical teams they collaborate with.
Introducing the AI Security learning path
The AI Security learning path is TryHackMe's direct response to the disconnect between AI realities and the security training available to individuals and teams. Built from those client conversations, structured around the OWASP LLM Top 10, and designed as a natural follow-on from Cybersecurity 101 (SEC1), it has the same building-block approach that's driven TryHackMe's success with foundational security training, applied to AI.
25 rooms across 5 modules, structured around the full OWASP LLM Top 10, Prompt injection but also the complete threat landscape from AI fundamentals through to secure AI systems, model security and data poisoning. It includes 7 challenge rooms built on a purpose-built AI platform, and there’s no prior AI knowledge required.
Three things make it different from anything else on the market.
First, learners practice on real AI systems. The same prompt gets different results every time, which is exactly the reality defenders face.
Second, the coverage is genuinely full-spectrum and beginner-friendly: offensive and defensive, covering every level of the stack.
And third, the supply chain module addresses the risk that clients raised unprompted: learners triage real model files, identifying backdoored tokenizers, poisoned checkpoints, and typosquat dependencies. It’s capturing the realities of AI risk in 2026.
AI Security skills for more than the security team
It's built for every level of the security team, from SOC analysts, to security engineers and architects, as well as the peripheral teams that are often overlooked in AI security conversations.
Max is clear on the urgency: "Every single engineer will need to understand this. It's no longer reserved for the senior security researcher. It's something that everybody that steps into the domain needs to understand." That includes developers using AI to write code, DevOps teams deploying models, and anyone building AI-powered features. "If you work for an organization as a developer or engineer and you're using an LLM on a daily basis, I would want to know I'm doing that in a secure way, and that it's not going to have greater implications downstream."
What your team walks away with
Vlad says that for practitioners, the goal is confidence. "It’s designed to help practitioners feel equipped, not scared or overwhelmed. Knowing that AI security isn’t some complex discipline reserved for people with machine learning PhDs. They can finish the path and say: I understand this, I can assess this, I know what to look for."
For the wider team, it's awareness that extends beyond the job. “Those outside the security team can leave feeling they're no longer behind, but also feeling ahead of peers who haven't done something like this.”
For managers, it's confidence about the team and a plan. "We can upskill the team, measure outcomes, and tell leadership we're not flying blind anymore.”
The AI Security learning path is available now, included with TryHackMe for Business. Talk to your account team about rolling it out to your team and colleagues.
Joanna Duffy