Blog/Inside CyberWatch AI

The Limits of Security Tools: What Technical Controls Cannot See

Filters, antivirus and MFA each solve a real problem, and none can judge whether a person should act on a message. Where the technical layer ends.

CyberWatch AI13 November 2026 · 6 min read
A rack of network equipment with cables in a server room

Here is an email. It has no attachment. Every link in it points to a domain with a clean reputation, or there are no links at all. It passes SPF, DKIM and DMARC, because the attacker registered the sending domain properly and configured it properly. The sending address differs from your supplier's by one character. The body says their bank account has changed, attaches the correct invoice number, quotes the correct amount, and asks that this month's payment go to the new details.

Now ask which of your security products should have stopped it.

The mail filter has nothing to score. The antivirus has no file. The endpoint agent has no process to inspect. Multi-factor authentication is irrelevant, because nobody is signing in to anything. Your zero trust architecture is working perfectly, and the accounts payable clerk is fully authorized to make the payment they are about to make.

This is not a failure of any of those tools. It is the shape of the boundary they share.

Every control answers a narrow question

Security tools get criticized for missing things they were never built to catch. It is worth being precise about what each one actually decides.

A mail filter decides whether a message resembles malicious mail: sender reputation, known-bad infrastructure, attachment type, link destination, volume patterns. It is good at this, and that is the problem. Filtering removes most malicious mail, so the population that survives is selected for its ability to survive. The message reaching a human inbox has already been through your best technical judgment and been passed.

Email authentication decides a different question, and it is frequently misunderstood. SPF, DKIM and DMARC confirm that a message was genuinely sent by whoever controls the sending domain and that the domain in the visible From address lines up. They are worth deploying. But they verify provenance, not honesty. An attacker who registers a domain one character away from your supplier's and sets it up correctly will pass every one of those checks, truthfully, because they really are the operator of that domain.

Multi-factor authentication decides whether the person signing in holds a second factor. This genuinely stops a huge class of attack: a password harvested from a fake login page is no longer enough on its own. CISA and the NCSC both push hard on it for good reason, and phishing-resistant forms close further gaps. None of that reaches an email which asks for a payment rather than a credential.

Zero trust decides what an authenticated identity may reach. It shrinks the blast radius of a compromise, which matters. It does not change whether a finance manager believes a plausible request and performs an action squarely inside their permissions.

Each control draws a clean line. Stack them up and they share one edge, and the edge is a person reading something and deciding what to do.

Why the gap is structural, not a product deficiency

You could argue that a sufficiently clever filter would catch the supplier email. In practice the attacks that matter most have been engineered specifically to look like business.

Consider what the attacker has removed. No malware, so nothing to detect. No credential request, so MFA has no part to play. No link, or a link to a legitimate file-sharing service that your filter cannot blocklist without breaking your business. No urgency so extreme it looks fake, just a deadline a real supplier might set. The email contains no technical indicator because the attack contains no technical component. It is a sentence that changes where money goes.

The same logic applies to callback phishing, where the email contains only a phone number and the attack happens on the call, and to attacks against MFA itself, where the lure is an approval prompt rather than a password field. In each case the technical component has been deliberately moved outside anything your tools inspect.

MITRE's catalogue lists these as techniques for exactly this reason. They are not edge cases. They are the mainstream of initial access.

The measurement gap is worse than the coverage gap

Here is the part that gets less attention. Every technical control reports on itself. Your filter tells you how many messages it quarantined. Your endpoint protection lists what it blocked. Your patch management shows coverage percentages. The security layer made of people reports almost nothing.

Most organizations can say, to a decimal place, how many threats their mail gateway stopped last month. Very few can say what proportion of their finance team would pay a changed bank account, how many people would report a payroll email asking them to confirm their details, or how long it would take from the first delivery of a real attack to the first internal report.

That asymmetry has a predictable effect on spending. What gets measured gets budget, and the human layer produces no dashboard by default, so it gets an annual training module and a hopeful assumption. A completion record is the nearest thing to a metric, and it measures attendance rather than behavior.

This is the gap worth closing, and it does not require believing that technology has failed. It requires treating the human layer as a control like any other: something you test deliberately, report honestly, and improve where the test finds a weakness. The practice of human risk management exists to do that, and NIST's cybersecurity framework has room for it in the functions that deal with awareness and training.

What actually helps

Four things change outcomes at this layer, and none of them is a lecture.

  • Process that does not depend on judgment. A rule that any change to supplier bank details is confirmed on a phone number you already held, never one in the email, removes the decision from the moment of pressure entirely. A written verification procedure beats a well-informed person having a bad morning.
  • A reporting route people use. If reporting is a forwarded email to a shared mailbox nobody answers, people stop. One step from the inbox, with an answer coming back, produces reports.
  • Practice on realistic material. People get better at recognizing what they have seen before. That means testing with the lures actually in circulation rather than the obvious ones.
  • Measurement you believe. Click rates by department, report rates, time to first report, and whether those numbers move after training. Without them you are managing a control with no telemetry.

Not sure about a message that reached your inbox? Paste it into CyberWatch AI Scan for a free second opinion before you act on it.

The honest version of the argument

Nobody should read this as a case for spending less on technology. The controls described above prevent an enormous volume of harm, and an organization that skipped them to buy awareness training would be making a much worse trade than the one it was trying to avoid. Patching, MFA, filtering and tested backups are not optional, and invoice fraud is not an argument against any of them.

The argument is narrower. The technical layer has a boundary, the boundary is where a person decides, and most organizations have no instrument pointed at that boundary at all. The phishing guide covers what arrives there; what is missing in most places is any honest answer to how their own people would respond when it does.

Frequently asked questions

Why do phishing emails get past the filter?

Filters remove the bulk of malicious mail, so what survives is selected for its ability to survive. An email with no attachment, no malicious link and a clean sending reputation gives a filter very little to act on.

Can an email pass SPF, DKIM and DMARC and still be an attack?

Yes. Those checks confirm that the message was genuinely sent by whoever controls the sending domain. If the attacker controls a lookalike domain they registered last week, their mail passes authentication honestly.

Does multi-factor authentication stop phishing?

It stops a stolen password from becoming an account takeover, which is a large win. It does not stop an attack that never asks for a password, such as a request to change supplier bank details.

Is antivirus useless against phishing?

No, but it is aimed at a different mechanism. Most payment fraud involves no file and no code, so there is nothing for endpoint protection to detect.

What about zero trust?

Zero trust narrows what an attacker can reach once inside, which is valuable. It does not change whether a person believes a convincing message and performs an action they are authorized to perform.

So should we spend less on technical controls?

No. Patching, MFA, filtering and backups do work that awareness cannot do. The argument is that they were never designed to answer whether a person should act on a message, and most organizations have no instrument that answers it.

What does cover the human layer?

Measurement and practice: realistic testing of how people respond, targeted training for the gaps the tests reveal, a reporting route people actually use, and verification steps built into the processes attackers target.

Which attacks are most exposed by this gap?

Those that rely entirely on persuasion: supplier bank-detail changes, executive impersonation, payroll redirection, gift card requests and callback scams that move the attack to a phone call.

Sources

  1. Phishing Guidance: Stopping the Attack Cycle at Phase One, US Cybersecurity and Infrastructure Security Agency
  2. Phishing: Spearphishing Link, Sub-technique T1566.002, MITRE ATT&CK
  3. Implementing Phishing-Resistant MFA, US Cybersecurity and Infrastructure Security Agency
  4. Multi-factor authentication for your corporate online services, UK National Cyber Security Centre
  5. Zero Trust Maturity Model, US Cybersecurity and Infrastructure Security Agency
  6. Business Email Compromise, FBI Internet Crime Complaint Center
  7. 10 Steps to Cyber Security, UK National Cyber Security Centre
  8. Cybersecurity Framework, NIST
Share this article
LinkedIn WhatsApp X
For organizations

Your people get these messages at work too.

CyberWatch AI sends your staff realistic practice attacks, trains the gaps it finds, and shows management exactly where the organization stands.