Security must evolve from a static training program into dynamic, AI-powered human and AI risk orchestration embedded across the organization.
The traditional model of security awareness aimed to get a message into people’s heads and hope it stayed there long enough to stop insecure actions. Organizations trained, tested, reported a completion rate to the auditor, and moved on. Yet, incidents kept happening year over year, highlighting that human behavior contributes to the vast majority of all cyber incidents (62% according to the Verizon DBIR 2026).
The sector quickly realized the problem was not only bad content, poorly delivered, but that cybersecurity professionals were asking the wrong question. The real question always was: how do I shape an environment in which my entire workforce is most likely to behave securely?
Today, this reframing matters even more because the workforce itself has changed. Organizations now deploy AI agents that work alongside employees — reading the same inboxes, touching the same systems, and making decisions at machine speed. Protecting this workforce is no longer a training problem; it is a behavior orchestration problem that must be answered twice: once for humans, and once for agents.
The traditional model of compliance-driven training established a vital baseline for security awareness. However, as the threat landscape has grown more sophisticated, sustainable security requires moving from knowing to doing. Two distinct factors drive this natural evolution:
Attention: Hermann Ebbinghaus mapped this with his forgetting curve: without reinforcement, the majority of newly learned material decays over time. Because a phishing email rarely arrives on training day, the future of security requires continuous, bite-sized reinforcement rather than a single annual event.
Behavioral Design: Awareness alone does not automatically dictate action. B.J. Fogg’s behavior model illustrates that action only happens when motivation, ability and a prompt converge at the exact same moment. While traditional training builds motivation, we can now also address ability and timing.
The next step is to build an environment where people are seamlessly empowered to act securely, making the secure path the easiest one by prompting employees exactly when it matters most. This is how we elevate standard awareness into a sustainable security culture where people behave securely instinctively.
The workforce itself has changed. For 20 years we taught people to hover over links and verify the sender. Now we are deploying AI agents that read those same emails, summarize those same documents, and draft those same replies at machine speed. AI agents, however, do not possess the same conscience and caution that govern human behavior.
Consider the EchoLeak class of attack: a prompt-injection technique in which a malicious instruction, hidden inside ordinary content an AI assistant processes, causes the agent to leak sensitive data without the user ever clicking anything. The agent simply reads what an attacker planted, treats it as an instruction, and acts.
An agent can be socially engineered, manipulated through the content it consumes, and given an instruction it should refuse but follows anyway. We are deploying AI agents that are deliberately probabilistic and non-deterministic by design because the value they create outweighs the imperfection.
But compromise for AI agents looks different from “normal” software. It looks more like a bad decision taken by a human, but at machine speed and with limited transparency. To detect this, you need to understand what the agent was told, what it reasoned, and why. That information is stored in an agent’s context, its prompt, and its reasoning trace. Agent security demands its own layer operating at runtime, with its own visibility into intent. The same behavioral standard as for humans, but different mechanics.
Organizations cannot orchestrate behavior at this scale manually. AI-generated spear-phishing and adaptive attacks do not wait for your next quarterly campaign, and a human SOC cannot detect rogue AI agents at machine speed.
Furthermore, the only intervention that actually changes human behavior must be personal. Delivering the right content, to the right person, at the right moment requires AI agents of your own — the only mechanism fast and granular enough to match a workforce that changes by the hour.
The agentic enterprise cannot treat human risk and agent security as separate problems; it must treat them as an integrated challenge managed from a central point of view. Compartmentalizing these risks into separate scores fails for two reasons:
A lack of coordinated intervention: Three scores in three tools is not a single coordinated intervention. That requires a unified risk model and an intervention plan for adaptive responses. Only a platform that sees the whole workforce can decide what to do about it.
A lack of culture building: Culture is built through what people repeatedly experience. Managing disconnected vendors means disjointed narratives. One coherent experience spanning prevention, intervention and remediation across attack simulation, collaboration security and agent security turns a security message into something people internalize.
One loop to measure security risk for the entire workforce is the right approach. An integrated risk score and intervention plan tells organizations where to look and what to do next, turning every human and AI agent into a trusted, defensible part of the business.
Effective defense today includes important design-time decisions alongside runtime governance for your entire digital workforce. Awareness alone is no longer never sufficient, and neither was the interception of threats without behavior change.
Because AI agents are now running business processes alongside humans while being similarly vulnerable to manipulation, organizations must establish runtime governance, tools, and processes that work seamlessly for both.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。