Skip to main content
BUSINESS-RESOURCE • 9 min read

How security teams are preparing for Agentic AI

Observations from a year of conversations with security leaders about defending the AI that their organisations are running.

In the space of about two years, enterprise AI has moved from chatbots that answer questions to agents that take actions. They read email, browse the web, call internal tools, hold credentials and hand work to other agents. Most of that change has happened outside the security team's line of sight.

In a 2026 survey of CISOs and security leaders by NeuralTrust, 72% said their organisation had implemented or was scaling AI agents, but only 29% reported comprehensive security controls around them. A Cloud Security Alliance study of 418 IT and security professionals found that 82% of enterprises had AI agents running in their environments that they didn't know about. Around two-thirds reported an agent-related incident in the past year that caused data loss or disruption. Among those, 61% involved data exposure, 43% operational disruption and 35% financial loss.

The security industry has started to formalise the problem. OWASP now publishes a dedicated Top 10 for Agentic Applications, separate from its list for LLMs, because agents fail in ways a chatbot can't. But a framework doesn't secure anything on its own. That job falls, of course, to security teams, but those teams often don't choose these deployments, and are evaluating their safety after the fact.

We've heard security leaders describe this tension time and again. One at an industrial equipment manufacturer framed the question most of their peers were asking: "what kind of risk is this introducing into our environment? And how can we detect and help mitigate that?" At a UK local council, a security professional put the underlying anxiety more plainly: "I'm just not confident it's as secure as the devs are telling me."

The industry is having the wrong conversation about AI

Read the industry press and you'd think the field had one story to tell: 'AI is coming for entry-level work, so upskill or become irrelevant.' Vendors have recently layered a second story on top: everything will be agentic and automatic. While both headlines get attention, neither provides constructive direction to SOC teams.

The change we're seeing when speaking to security teams is more nuanced. AI won't transform the field uniformly. It will move at different speeds in different SOCs, industries and verticals, and it will reach every seniority level of a team.

We're noticing that work is moving down the career ladder, with tools making execution easier for less experienced team members. Detection engineering is a good example. It was long treated as principal-level work, but it's fast becoming a mid-level task as AI takes on more of the heavy lifting. Triage is changing as well. AI will change senior roles as much as it will change junior ones. AI security knowledge can't stay a specialism either. Every engineer who works with AI needs to understand it, not just senior security researchers.

This is, of course, an educated bet. No one can predict the future. But a cyber security training company has two choices: it can wait for the market to settle and build courses six months behind, or it can take a view on what good will look like and help teams get there through the uncertainty. We've chosen the second, building the proof one practical, hands-on path at a time.

From what AI says to what it does

The first wave of AI risk was about data leaving the organisation. For many teams, it still is. A security professional at a European automotive finance company named the problem as follows: "users are using AI in a non-compliant way like uploading company internal information into public AI tools." For most organisations, that's where a structured approach to AI security began: with acceptable-use policies, DLP rules and a lot of chasing.

But the conversation has evolved alongside the systems. When AI can act, a mistake is itself an action. A security professional at a US services company described the scenario that worried them most: "if we had an AI process go rogue and cause us to lose some data or corrupt the system, that would be pretty significant." Another, at a global medical technology company, described an agent with internet access scraping a website that carries a hidden instruction, and wanted to know what that agent could do and how to put guardrails around it.

The common theme was visibility, or the lack of it. "We do not have visibility into what the agents can do within our system and within our infrastructure," a security leader at a global IT services company told us. Others said the same. A US healthcare company didn't yet know how to monitor for prompt injection, while a health insurer wanted to understand what its agents are doing and what they can access, without restricting them so far that they stop being useful.

None of this was new to us. Before building our AI Security path, we spoke to security leaders across healthcare, finance, automotive, manufacturing and insurance to ask what they actually needed. The answers were remarkably consistent. AI was already in production handling confidential data, and security teams were being asked to protect systems they didn't understand end to end. Supply chain risk came up unprompted in almost every conversation. And the training on offer was either too expensive, too advanced for most of the team, or too passive to build real skills. So we built what they asked for.

The path came straight out of those conversations. That meant learners practising on real AI systems rather than simulations, because a simulated model always gives the same answer and a real one doesn't. It meant offensive and defensive coverage pitched so the whole team could take part, not just senior specialists. And it meant a supply chain module built around the risk leaders kept raising unprompted. Our new Agentic AI Security path is the next chapter of the same story, taking the same approach to learning through exposure to real AI agents that act on their own, on a live multi-agent system that learners can hijack, trace and secure.

Everyone is learning on the fly

If there was one admission we heard more than any other, it was that security teams are learning about AI at the same time as they're being asked to secure it.

A security leader at a US energy utility was candid about where the AI security gap starts. "For our analysts there's like a general level of awareness," they said. "I don't think the majority of at least our tier one analysts really know what [AI security] means, or really know the specifics around it." A global retailer made the same case from the ground up, asking for the basics, like what RAG and MCP are, so its security staff can follow what's being built and protect it.

The gap isn't limited to junior analysts. Even the people closest to AI are learning on the job. At a European insurance group, the teams building AI projects were working through security training while delivering those projects, and a security professional there told us those teams "have to learn a lot of things on the fly." And the need is immediate: at a UK supermarket chain, a security professional said that understanding "how the attacks using the generative AI works" is "pretty important for the team right now."

The training market hasn't caught up with that need. Most AI security content falls into one of two camps. The first is high-level offensive material on how to pentest AI, written for experienced red teamers. The second is high-level governance that restates frameworks without giving space for realistic practice. There isn't much to create a practical baseline, or to prepare a blue team to defend an agent that has just been hijacked.

Passive formats hit a ceiling quickly. Understanding prompt injection from a lecture is very different from trying to extract a system prompt from a real model with guardrails in place. A security professional at a global industrial automation manufacturer summed up the shortfall: "the systems, they can be broken, you know, quite easily. And I don't think there's enough content out there that shows how to actually protect."

Defenders want to hold the attacker's tools

When security leaders described the training they wanted, they rarely picked a side. They wanted red and blue together, because both functions defend the same systems.

An industrial automation manufacturer wanted its red team pentesting AI systems and its blue team monitoring them. A security professional at a US HR and payroll software company asked for both in one breath: "what kind of guardrails can security secops put around our AI tooling? And then also how can we test the security controls around those?"

Organisations want testing they can trust. A security professional at a US medical device maker worried about how easily systems handling patient data could be compromised, and was blunt about the checks already in place: "We have pen testers, but to me, they're just box checkers, you know, they just check the box." They wanted to ensure that their red teams genuinely understood how AI breaks, and their blue teams had hands-on experience defending.

Teaching AI as an attack surface

AI should be taught the way security teams learn every other system: understand how it's built, break it, watch what happens and fix it. That's why we treat AI as an attack surface, not a theoretical subject.

We're focused on producing realistic, hands-on AI security labs, particularly where teams are currently underserved: AI security that's accessible at every level and gives every practitioner a real baseline, alongside agentic defence that almost no one else covers.

Our AI Security path offers the strongest entry-level AI security content on the market. Its 25 hands-on labs, structured around the OWASP Top 10 for LLMs, take learners from AI fundamentals through threat modelling, prompt injection, supply chain attacks and RAG poisoning, attacking and then defending at each step. Every lab runs on live models that can answer the same prompt differently each time, so learners build the judgement production systems demand. Teams are already using it alongside live AI projects. An insurance group's security professional called it "a real game changer" for their AI teams.

The Agentic AI Security path picks up where that foundation ends, at the point where AI starts acting. Learners work on a live multi-agent platform where an orchestrator and sub-agents have their own models, identities and tools, and talk over a message queue that can be poisoned or forged. An inspector view then traces every tool call, message and identity, offering the visibility security leaders told us they're missing.

Five modules cover the full OWASP Top 10 for Agentic Applications (2026), ASI01 to ASI10, in the order an attack unfolds: getting in, abusing tools, memory and authority, then full compromise. Every module covers attack and defence. Learners exploit an agent, then trace the damage and build the control that would have stopped it.

We've taken a different approach to teaching AI security than the rest of the market. Most AI security courses stop at prompt injection, offer attack-only tradecraft or summarise frameworks. The Agentic AI Security path is structured to take learners through all ten OWASP agentic threat categories, as attacker and defender, on a live multi-agent system.

The challenge rooms are built on real incidents, including EchoLeak, the Gemini memory attack, AutoGPT remote code execution, the GitHub MCP exploit and the Replit meltdown. Like all our content, it's self-paced and browser-based. With the AI Security path and the AI1 certification, it forms one journey from foundations to multi-agent defence.

Readiness before the incident

None of the security leaders we spoke to wanted to stop their teams using AI. They wanted confidence in their people: teams who can see what AI systems are doing, spot when one has been turned against them and respond without waiting for a specialist. We designed the path against their stated need to upskill the whole team, measure progress and be able to demonstrate that readiness.

AI won't impact every team at the same pace, which is why training needs to mark a path now, instead of catching up later. Some teams will be defending agents in production next quarter, while others won't meet one for another year.

Analysts may see their day-to-day reshaped first, and senior specialists will find their roles shifting in different ways. Waiting for the dust to settle would be complacency. We think a training company owes its learners a clear path through that uncertainty. In practice, that means listening to the people defending these systems, building practice on real AI, shipping it while a threat is still taking shape, and being open about what we think good will look like. That's the commitment behind what we've built so far, and it's how we'll keep building as threats keep changing. The most resilient teams won't have read the most about AI risk. They'll have already broken a live agent, traced the damage and built the fix.

Explore the new Agentic AI Security path here.

Read more insights on how security teams are bringing AI into the SOC in our new report, the Hard Part is Still Human.

Joanna Duffy
Nov 16, 2026

Recommended

Get more insights, news, and assorted awesomeness around cyber training.

Business • 10 min read

How Banks Can Prioritise Cybersecurity Exercises Using NIST CSF 2.0

The NIST Cybersecurity Framework (CSF) 2.0 offers a useful way to make that decision. By connecting cybersecurity outcomes to existing evidence, banks can identify capability gaps, prioritise exercises and demonstrate whether their teams are improving.

Business • 9 min read

How to Run a Ransomware Tabletop Exercise for Banks

A ransomware tabletop exercise gives banks an opportunity to work through those situations before a real incident. By introducing realistic developments and examining how participants respond, security leaders can identify gaps in their procedures, decision-making and recovery arrangements.

Careers • 12 min read

SOC Analyst Certification Comparison 2027: Which One Should You Do First?

In this guide, we'll compare five popular SOC and blue team certifications for 2027, covering their exam formats, difficulty, costs, and career relevance so you can decide which one to take first.

Join over 640 organisations upskilling their
workforce with TryHackMe