As AI agents start working alongside employees, classic protection reaches its limits. At it-sa, KnowBe4’s Dr. Martin Krämer explains why people, email filters and agent monitoring must be secured together.
The rapid rise of artificial intelligence is reshaping the enterprise at unprecedented speed. As AI agents increasingly work alongside human staff, the boundaries of IT security are shifting. At the it-sa security fair in Nuremberg, Dr. Martin Krämer, CISO Advisor for Europe and the Middle East at security specialist KnowBe4, explained why traditional safeguards fail against autonomously acting software.
For years, the debate on AI in business has centered on efficiency. Now organizations must secure that promised productivity in a highly dynamic threat landscape. Krämer’s starting point is that the modern workforce no longer consists of people alone, but also of autonomous AI agents. To run companies effectively in the AI era and make cybersecurity a business enabler, the digital workforce must be empowered, protected and integrated as a whole. “It is really about ensuring productivity,” Krämer says.
Critical thinking is quietly eroding
The core problem, he argues, lies not only in a growing number of technical entry points but in a psychological effect. Automation and uncritical trust in convincingly worded AI output cause the habit of questioning to wither in daily work. The brain outsources cognitive effort to algorithms, a risky path when it comes to spotting manipulation.
AI has also given social engineering a new dimension. Classic email fraud used to give itself away through poor grammar or conspicuous attachments. Today, attackers work across many parallel channels: deepfakes in video calls, cloned voices on the phone (vishing) and forged messages in Microsoft Teams or Slack. Experts may debate how convincing AI fakes really are, but the sheer volume of attacks has risen sharply. Even if quality were equal, Krämer says, quantity itself becomes a quality feature.
Added to this is the plausibility effect. AI-generated images that stage an emotional emergency or stoke political tension look entirely believable. Krämer cites fabricated flood images from the US showing young girls in a boat with their dogs, captioned to claim the government had abandoned them. Employees therefore need to be trained again to ask four questions: Why does this information exist? Who created it? What is their intention? And what reaction is it meant to provoke?
Three pillars for a security culture
Ticking compliance boxes is no longer enough, Krämer says. KnowBe4 describes a platform built on three interlocking pillars.
The first is security awareness training combined with attack simulation. Instead of dry theory, the company relies on narrative formats such as “Inside Man,” an award-winning drama series now in its seventh season. According to Krämer, some customers unlock the series after mandatory training and reach training rates close to 100 percent.
The second pillar is communications security, above all cloud email security, now being extended to Teams and Slack. Natural language processing and machine learning learn typical communication patterns to detect business email compromise or messages whose only payload is a phone number for voice phishing. Krämer stresses that this is not about monitoring staff. By his rough estimate, a secure email gateway filters about 90 percent of threats, and integrated cloud email security catches another 7 to 9 percent. The remaining 1 percent, the most critical, needs trained employees as a human in the loop.
When AI agents become insider threats
The third pillar covers managing and monitoring AI agents, and it changes the most. Because agents generate their steps at runtime, there are no static signatures to check beforehand. Responsibility also stays with people: if an agent leaks data or halts operations, executives and the employees who built it remain accountable, not the machine. Flawed system prompts or manipulative inputs can make an agent lose control and act like a digital insider.
A deployed model cannot be trained like an employee, so runtime monitoring is the only option, with a kill switch as the last resort. The governance process is clear, Krämer says: the security team shuts the agent down and informs the responsible employee, who corrects the instructions. After review, the agent goes back online. KnowBe4 is building a tool to monitor AI with AI; it is in tech preview, with general availability expected early next year.
Culture decides
Shadow AI, phishing waves or autonomous agents: in the end, culture decides whether a security architecture works. Technology and policies provide guardrails, but lasting resilience comes only from combining advanced filtering with a critically thinking workforce. Companies that treat this shift as a continuous improvement process can turn the human risk factor into an effective line of defense. According to Krämer, more than 60 percent of security incidents involve the human factor. KnowBe4 says its AI-related annual recurring revenue grew by more than 35 percent in 2026 and that more than 70,000 customers use its platform.

Dr. Dietmar Müller is an experienced IT journalist and editor-in-chief of Beyond Buzzwords. He has decades of expertise in the German and international IT markets, as well as extensive editorial experience at renowned publishing and media companies such as Springer, IDG, and CBS Interactive. Whether writing cross-media technical articles, conducting in-depth analyses, or conducting insightful interviews—whether at his desk or in direct dialogue with industry leaders—Dr. Dietmar Müller is right at home in the world of enterprise IT. His key areas of focus include enterprise IT and OT architectures, the cloud, ERP, MES, data centers, cybersecurity and compliance, industrial automation, and AI governance, among others.