The Future of Human Cybersecurity: Eight Predictions, Clearly Labelled
Where human-layer security is heading as AI makes social engineering cheap and convincing. Eight predictions, each with what would prove it wrong.

Forecasting in security is mostly a way of restating the present with more confidence. So this article separates two things that usually get mixed together: what is observable now, and what I think follows from it. Everything in the second half is labelled as a prediction, and each one comes with the condition that would show it to be wrong.
What is observable now
Four things are not in dispute.
Attacks that persuade a person need no vulnerability. There is nothing to patch in a message asking your accounts team to update a supplier's bank account. The FBI's annual crime reports keep business email compromise near the top by reported loss, and the mechanism involves no exploit at all.
The quality tells are disappearing. The NCSC's assessment of AI's near-term impact on the cyber threat expects more convincing social engineering, produced faster and in greater volume. Poor spelling and clumsy phrasing were never security controls. They were artifacts of attacker effort, and that effort just became cheap.
Attacks are moving off the channel you protect. A QR code moves the link to a personal phone, past the filtering on a work computer. A callback number moves the attack to a phone conversation that no mail gateway can inspect. Chat apps and SMS arrive with no filter at all. We see all of these in what people report, and it is a deliberate shift rather than an accident.
Most organizations still cannot measure the human layer. They can report patch coverage, MFA enrollment and filtering volumes to a decimal place, and when asked how their people would respond to a convincing attack, they produce a training completion percentage.
Eight predictions
Prediction 1: recognition stops being the primary defense
Teaching people to spot phishing by its appearance has maybe a few more years as a front-line strategy. As the writing gets better and the pretexts get more specific, visual tells thin out to the ones that are structural rather than stylistic: a link that does not go where it says, a request that bypasses a normal process, a payment detail that changed.
What replaces it is verification built into the process. A rule that any change to supplier bank details is confirmed on a number you already held, applied regardless of how convincing the email is, does not degrade as attacks improve.
What would prove this wrong: detection tooling getting good enough, at the point of delivery, that convincing attacks rarely reach people at all. Possible, and I would not plan around it.
Prediction 2: human risk becomes a reported metric
Patch coverage and MFA enrollment are reported to boards, insurers and procurement teams as a matter of routine. Human risk is not, mostly because there has been nothing to report.
I expect that to change within a few years: a measured figure for how an organization's people respond to realistic attacks, with a trend, appearing in the same packs as the technical controls. Insurance questionnaires and supplier security reviews will be the forcing function rather than any voluntary enthusiasm.
What would prove this wrong: the measurement proving too easy to game. If organizations can produce a flattering number by running easy tests, the metric will be discounted quickly and deservedly.
Prediction 3: the annual training module dies, slowly
It will survive wherever a compliance regime names it, which is a lot of places. As a serious intervention it is already finished, and its replacement is visible: short lessons, aimed at what a specific person actually got wrong, delivered close to when they got it wrong.
The associated change is that completion stops being treated as an outcome. The outcome worth reporting is behavior before and after, which is harder to produce and much harder to argue with.
What would prove this wrong: regulation that specifies annual training in a way that makes continuous programs harder to evidence than the module they replaced.
Prediction 4: simulations get written per organization
A static template library ages. The lures in circulation shift every few months, and a test using last year's pretexts measures last year's problem.
The direction is toward simulations generated for a specific organization from a pattern currently in use, with a human approving each one before it sends. We already build this way: an administrator can start from a reviewed list of attack patterns circulating now, or describe one, and the email is written for their organization, with the warning signs recorded so the page behind the link can name that email's own tells. The approval step is not going away, and should not.
What would prove this wrong: organizations deciding that generated lures are too unpredictable to authorize, and preferring a reviewed library they fully control. This is a reasonable position and some will take it.
Prediction 5: reporting speed becomes the headline number
Click rate is the metric this field was built on. I think time to first report overtakes it, because it is the number that determines whether a security team can act before the rest of the organization opens the same email.
It also has a property click rate lacks: it rewards a culture where people speak up quickly, which is the behavior that actually helps during an incident.
What would prove this wrong: automated detection and containment getting fast enough that human reporting stops being on the critical path for most incidents.
Prediction 6: AI assistants become a human-layer attack surface
This is the prediction I am most confident about and the one organizations are least prepared for.
An assistant that reads your email, your documents and web pages on your behalf is reading text controlled by whoever wrote it. OWASP ranks prompt injection first among risks to applications built on language models, and NIST's adversarial machine learning taxonomy treats it as a core attack class. The practical version is an email whose hidden instructions are aimed at your assistant rather than at you, attempting to act with your authority.
The defenses are architectural rather than educational: treat analyzed content as untrusted, keep instructions on the server rather than accepting them from the client, and require a human approval for anything that changes state. That is how our own engine is built, where every change an agent wants to make is queued as a proposal for an administrator to approve or decline, so an instruction hidden in a phishing email can at most generate a suggestion somebody reads.
What would prove this wrong: a genuine technical solution to prompt injection. Several people are working on it, and so far nothing has held.
Prediction 7: the human-layer attack surface widens past email
Email gets the attention because it is where the measurement tooling grew up. The attacks are already distributed across channels: cloned voices on phone calls, chat apps, SMS, QR codes, meeting invitations, and increasingly a sequence that crosses several of them in one operation.
Testing and training will have to follow, which is awkward because the channels nobody controls are exactly the ones where simulation is hardest to do ethically. Expect the measurement to lag the threat here by some years.
What would prove this wrong: platform-level controls on messaging and voice maturing faster than expected, pushing attackers back toward channels organizations can see.
Prediction 8: attempts to standardize human risk scores will overreach
Once human risk is reported, somebody will try to make it comparable between organizations, and a benchmark will be published.
It will be unreliable, because a score of this kind depends on how hard the tests were, what mix of roles the organization has, and how reporting is configured. Two companies with identical scores can be in very different positions. The useful comparison is against your own previous months, and a benchmark will be cited anyway.
What would prove this wrong: an agreed difficulty standard for simulated lures, rigorous enough that cross-organization comparison means something. NIST's Phish Scale is a step in that direction, so this is not impossible.
What I am not predicting
Three things that get forecast regularly and that I think are wrong.
That people become unphishable. No training program produces this, and designing around it is how organizations end up blaming individuals for structural failures. The achievable goal is that fewer people fall for common lures, more people report quickly, and the processes attackers target stop depending on anyone's vigilance.
That AI detection wins. Detection and generation improve together. Treating the arms race as winnable leads to under-investment in the things that do not depend on winning it, such as out-of-band verification and phishing-resistant authentication.
That the human layer gets absorbed into the technical layer. Every few years something is proposed that will make the question moot. The question has outlived all of them, because it is not a technology gap. Somebody reads something and decides.
The useful version of all this. You cannot plan for eight predictions. You can make verification a property of your payment process, make reporting one step, and start measuring how your people respond so you notice when any of this changes.
What follows from it
If the predictions above are roughly right, the organizations in the best position in five years will not be the ones with the most training hours logged. They will be the ones that made four unglamorous changes: verification built into the processes that move money, a reporting route that takes one step and gets answered, continuous measurement of how people actually respond, and an architecture that assumes content reaching an AI system was written by an adversary.
None of those depend on predicting the threat correctly, which is the main reason to prefer them.
For where to start today, human risk management covers the discipline, AI-written phishing covers what is already arriving, and the phishing guide covers the attacks these predictions are about. Anyone can check a specific message for free with CyberWatch AI Scan.
Frequently asked questions
Will people stop being the main target?
There is no sign of it. Attacks that persuade a person require no vulnerability, survive patching, and cost very little to attempt, which is a combination that keeps working.
Does AI mean we should stop teaching people to spot phishing?
It means spotting should stop being the primary defense. Recognition still helps, but as the writing quality of attacks rises, verification steps built into processes matter more than visual tells.
What replaces the annual training module?
Continuous, short, targeted learning driven by measurement: a lesson aimed at what a person actually got wrong, delivered close to when they got it wrong.
Will AI detection keep up with AI-generated attacks?
Partially, and not decisively. Both sides improve, which is a reason to reduce how much depends on detection rather than to bet on winning the race.
Are AI assistants a new human-layer risk?
Yes. An assistant that reads email, documents or web pages on your behalf is reading attacker-controlled text, which makes prompt injection a path to acting on your authority.
Will human risk scores become standardized?
Attempts are likely, and cross-organization comparison will remain unreliable because scores depend on test difficulty, role mix and reporting setup. Comparison against your own trend stays the honest use.
Which attacks will grow fastest?
Those that move off email entirely: QR codes to a personal phone, callback numbers, chat apps and voice. Each sidesteps the filtering that protects a work inbox.
What single change helps most over the next few years?
Making verification a property of the process rather than a property of the person, particularly for any change to where money goes.
Is any of this certain?
No. Everything below the observations in this article is a prediction, and each one is stated with what would show it to be wrong.
Sources
- The near-term impact of AI on the cyber threat, UK National Cyber Security Centre
- AI Risk Management Framework, NIST
- AI 100-2 E2025, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations, NIST Computer Security Resource Center
- LLM01:2025 Prompt Injection, OWASP Gen AI Security Project
- Annual Reports, FBI Internet Crime Complaint Center
- ENISA Threat Landscape 2025, European Union Agency for Cybersecurity
- Internet Organised Crime Threat Assessment (IOCTA), Europol
- Implementing Phishing-Resistant MFA, US Cybersecurity and Infrastructure Security Agency
- Cyber Security Toolkit for Boards, UK National Cyber Security Centre


