AI phishing training is shifting from completion to behavior

At a glance

What this is: This research argues that phishing awareness has moved beyond completion and click tracking, because AI-generated attacks now imitate normal business workflows and make traditional red-flag training less reliable.

Why it matters: IAM and security leaders need behaviour-based measures because participation metrics can look healthy while employees still engage with realistic phishing, vendor fraud, and impersonation attacks.

Context

Security awareness programmes often measure whether people completed training or clicked a simulation, but those signals do not show whether behaviour changed when a real attack arrived. That gap matters most when phishing is personalised, role-aware, and designed to look like ordinary work rather than an obvious scam.

For identity and access teams, this is not just a training problem. Human decision-making remains a control point in IAM and broader security programmes, and the article’s core claim is that modern phishing now exploits that control point faster than annual or quarterly awareness cycles can adapt.

Key questions

Q: How do organisations measure whether phishing training is actually changing user behaviour?

A: Organisations should track behaviour-based signals, not just completion rates. Useful measures include simulation interactions, reporting activity, follow-up training outcomes, and trend lines in user risk scoring. If those indicators improve over time, the programme is influencing decisions under pressure. If they do not, the training may be compliant on paper but ineffective in practice.

Q: Why do click rates and completion rates give false confidence in security awareness programs?

A: Because they measure participation in a programme, not resistance during an attack. An employee can finish training, pass a simulation, and still fall for a role-based impersonation that looks like normal work. The metric gap matters most when phishing is personalised, because the attack succeeds by blending into legitimate business communication.

Q: What breaks when phishing no longer contains obvious red flags?

A: Controls that depend on spelling errors, poor grammar, or generic phrasing lose reliability. Teams that still use message appearance as a primary trust signal will miss more malicious emails, which means detection has to move toward provenance, behaviour, and downstream identity impact.

Q: How do teams reduce human error in vendor fraud and executive impersonation attacks?

A: Use role-specific simulations and immediate coaching for the teams most often targeted, especially finance, procurement, and executive support. These roles need scenarios based on real request patterns and trust relationships, because the attack is usually a business-process impersonation rather than a generic malicious email. Behavioural tuning is more effective than one-size-fits-all awareness.

Technical breakdown

Why classic phishing training no longer matches modern attacks

Classic phishing awareness was built around visual tells such as odd sender addresses, spelling mistakes, and suspicious links. That model assumes the attacker is trying to look fake. AI-generated phishing changes the game by mirroring real roles, business processes, and communication patterns, so the message can look routine to the target. When a finance employee sees an invoice that resembles normal vendor traffic, the training cue disappears. The failure is not just technical filtering, but the mismatch between how people are taught to detect threats and how attackers now construct them.

Practical implication: update training content to reflect realistic business workflows, not legacy phishing stereotypes.

Behavioural measures versus participation metrics

Completion rates and click rates tell you whether employees attended training or reacted to a test, but they do not prove resilience during a live attack. Behaviour-based measurement looks at what people actually do under pressure: whether they report suspicious messages, whether they escalate fast enough, and whether they enter credentials or approve requests they should have questioned. That shift matters because security awareness is only useful if it changes decisions in the moment, not if it only produces audit-friendly attendance records.

Practical implication: measure reporting, escalation, and risky-action rates instead of relying on course completion alone.

Just-in-time coaching and agentic delivery models

The article describes instant coaching delivered the moment a mistake happens, which is materially different from annual awareness cycles. The learning event is tied to the exact signal the employee missed, such as an unusual request pattern or a subtle impersonation cue. It also notes an autonomous delivery model where the training system manages simulation creation, delivery, and adaptation. That shifts awareness from a periodic programme into a responsive control loop that can scale across a large workforce without depending on manual template management.

Practical implication: align awareness programmes to real-time feedback loops so coaching follows exposure, not calendar cycles.

Breaches seen in the wild

• Mailchimp breach 2022: Attackers socially engineered Mailchimp staff, used a support tool to export 102 customer lists and exposed customer API keys for phishing.

Read and download The State of NHI & AI Agent Breach Report 2026, covering 150+ breaches impacting Non-Human Identities including AI Agents.

NHI Mgmt Group analysis

Behavioural validation is now the only defensible measure of phishing resilience: completion data proves participation, not resistance to attack. When phishing blends into legitimate work, the question changes from whether employees were trained to whether they changed decisions under pressure. That is an IAM and security governance problem because the control objective is human action at the point of trust, not course administration.

AI-generated phishing has collapsed the old 'fishiness' model: training built around obvious red flags was designed for attacks that looked fraudulent on sight. Personalised messages tied to real roles and workflows remove those cues, so the programme must now measure recognition of context, not just detection of spelling mistakes. The practitioner implication is that awareness content should be built around business-process impersonation, not generic scam tropes.

Human risk management is becoming a live control plane rather than a compliance artefact: once behaviour is measured continuously, leaders can see where risk clusters around specific teams and task types. That turns awareness from a box-ticking exercise into a governance discipline that can inform targeted intervention for finance, procurement, executive support, and other high-trust roles.

Instant coaching creates a more credible learning loop than quarterly phishing tests: the article points to a model where feedback arrives at the moment of error, which is when recall and correction are strongest. The broader lesson is that resilience improves when training is linked to live behaviour and role context, not when it is separated from the incident by months of delay.

Behaviour drift in high-trust roles: vendor fraud and executive impersonation exploit business familiarity, not technical weakness. That means programmes need to treat finance, procurement, and executive support as differentiated risk populations, with measurement and coaching tuned to the way those roles are actually targeted.

What this signals

Behaviour-based awareness will displace participation-based reporting as the meaningful control signal. Security leaders should expect more programmes to move from annual training completion toward continuous measurement of reporting, escalation, and risky-action rates. That shift is important because phishing success is now driven by context and timing, not just by whether the target has seen a slide deck.

Role-specific coaching will become the practical test of awareness maturity. Teams that can tailor simulations to finance, procurement, and executive support will expose more of the actual exposure surface than teams running the same generic test for everyone. The programme signal is simple: if the training does not reflect the way your business is actually targeted, it is measuring the wrong thing.

For practitioners

• Replace completion metrics with behaviour metrics Track reporting rate, escalation speed, credential entry, and risky approval behaviour instead of treating training attendance as proof of resilience.
• Build simulations from real attack patterns Use intercepted or observed phishing themes that reflect the roles, relationships, and workflows in your environment rather than generic templates.
• Deliver coaching at the moment of error Route users to immediate, personalised feedback when they click, submit data, or ignore a warning so the correction is tied to the exact mistake.
• Segment high-risk populations Measure finance, procurement, and executive support separately because those roles are more likely to be targeted with business email compromise and vendor fraud.
• Automate the awareness lifecycle Use an operating model that can create simulations, adapt content, and maintain year-round coverage without relying on one person to manage quarterly campaigns manually.

Key takeaways

• AI-driven phishing now behaves like ordinary business traffic, which makes legacy red-flag training less reliable as a resilience measure.
• Completion and click metrics can confirm participation, but they do not show whether people make better decisions when a realistic attack arrives.
• Security teams should shift awareness programmes toward behavioural metrics, role-specific simulations, and immediate coaching at the moment of error.

Key terms

• Human Risk Management: The practice of managing how people interact with security controls, especially under pressure, distraction, or deception. It combines training, policy, and friction management so identity systems are still usable enough that users do not bypass them in day-to-day work.
• Vendor Email Compromise: Vendor email compromise is a form of impersonation that targets supplier, contractor, or partner relationships. Attackers exploit routine vendor communication patterns to request payment changes, invoice redirection, or other sensitive actions, so identity and process verification must extend beyond internal users.
• Behaviour Metrics: Behaviour metrics are measures that show how people actually respond to risk, such as repeat clicks, follow-up susceptibility, or improvement after coaching. They are more useful than vanity metrics because they indicate whether the programme is reducing exposure instead of merely documenting activity.
• Role-Specific Simulation: Role-specific simulation is phishing testing built around the actual communications, workflows, and trust relationships associated with a user’s job function. It is more effective than generic testing because the target sees the kind of message they are most likely to receive in daily work.

Deepen your knowledge

NHI governance, agentic AI identity, and machine identity lifecycle are core topics in our NHI Foundation Level course, the industry's only accredited NHI security programme. If you are responsible for identity security strategy or NHI governance in your organisation, it is worth exploring.