Why We Built CyberWatch AI
Scams stopped looking like scams, and nobody had a fast way to ask whether a message was real. The three problems CyberWatch AI was built to solve.

Scams stopped looking like scams. That sentence is the whole origin of this product, and it is worth taking literally.
The misspelled prince, the demand for a wire transfer in broken English, the web address with four hyphens in it: those were real and they are now the exception. What arrives instead has clean grammar, the right logo, a sender name you half recognize, and a request urgent enough that checking it feels like an insult to whoever sent it. The UK's national cyber agency has published the observation that telling people to avoid clicking bad links still is not working. They are right, and the reason is that the advice assumes the bad link looks bad.
The first problem: nobody had a fast way to ask
The pattern we kept seeing was not carelessness. It was people with reasonable judgment making a decision in eleven seconds, with no way to check, about a message designed by someone who had spent considerably longer than eleven seconds on it.
Ask a friend in IT and you wait. Ask the bank and you queue. Search the sender's address and you find nothing, because the domain was registered two days ago. And underneath all of that is the quieter problem: asking at all feels foolish. Nobody wants to forward a message to a colleague and be told it was obviously fine.
So the first thing we built was a way to ask the question privately and get an answer in seconds. Paste the link, the message or a screenshot, and get a risk score, the specific warning signs, and what to do next. Free, with no account needed, because the moment someone needs it is not the moment to meet a sign-up form.
Two things about that tool came from watching how people actually use it. It has to read the message rather than look it up, because a scam running in Accra or Manila in a language the blocklists have not caught up with is still a scam. And it has to admit uncertainty, because a tool that is confidently wrong once teaches somebody to stop checking.
The second problem: inside organizations, the same gap at scale
The same thing happens in companies, with a bigger invoice attached.
Money goes to the perimeter, and the perimeter is not where the attack arrives. A mail filter removes a great deal, which is precisely why anything reaching an inbox is from the small population of messages engineered to survive filtering. By the time a person is reading it, the control you were relying on has already been passed.
And the pattern of the attack has shifted away from anything a technical control recognizes. Business email compromise contains no malware, no exploit and often no link. It is a supplier writing to say their bank account has changed. The letterhead is right because it was copied, the reference number is right because it was copied, and the only thing that is different is the account number. Europol's organized crime threat assessments and the FBI's annual complaint reports both keep describing fraud whose entire mechanism is persuasion.
Which leaves one question that matters: would our people recognize that? And almost nobody could answer it.
There is a second-order effect worth naming. When nobody can answer the question, the organization cannot tell the difference between being well defended and being lucky. Two years without a successful attack might mean the finance team is sharp, or it might mean nobody has aimed a properly researched email at them yet. Those two situations look identical from the inside and call for completely different responses.
The third problem: assumption dressed as evidence
Where organizations did try to answer it, the answer was usually a completion record.
Everyone sits a module once a year. The platform records that 94 percent finished it. That number goes in a report, and the report implies the human layer is handled. But the completion record measures attendance. It does not tell you whether your operations team clicks a parcel redelivery notice, whether anyone would report a payroll email that asked them to confirm their bank details, or how long it would take from the first delivery to the first report. Those are different questions, and they have answers.
NIST rewrote its own guidance on this in SP 800-50 Revision 1, moving from awareness and training as an activity to a learning program with measurable outcomes. The instinct is correct. The missing piece, for most organizations, is an instrument.
So the second thing we built was that instrument. Send realistic phishing to your own people, with an administrator's authorization and no real credential ever collected. Watch who opens, who clicks, who reports and how fast. Send a three-minute lesson to the people the test identified, on the exact tell they missed. Then measure whether the next round is better. The loop runs all year instead of once.
Why the free tool and the paid product are the same argument
These look like two businesses. They are one position stated twice.
An individual needs an answer to "is this real?" in the thirty seconds they have. An organization needs an answer to "how would my people respond?" before the day it matters. Both are the same layer of the problem, the layer where a human being reads something and decides. Both have been chronically underserved compared with the layer where packets move.
It also explains the commitments that look commercially odd. Scanning is free with no daily limit, nothing scanned is used to train models, the content itself is discarded after analysis, and in an organization an administrator never sees a check an employee chose not to report. A tool people are wary of is a tool people stop using, and a human-layer product whose users avoid it has failed at the only thing it does.
Checking something is not an admission. If a message is making you hesitate, that hesitation is the useful signal. Paste it into CyberWatch AI Scan and decide afterwards.
What we are not claiming
None of this makes technical controls optional. Multi-factor authentication stops credential theft from becoming account takeover. Patching closes the holes that do not need a person at all. Filtering removes the volume that would otherwise bury your staff. Human-layer security sits on top of those, and if anyone offers it as a substitute, that is a gap being sold as a product.
Nor are we claiming that people can be made perfect. The honest goal is narrower. Fewer people fall for the common lures, more people report quickly, the team knows where its own weak spots are rather than guessing, and the first report arrives in minutes instead of after the payment has cleared. Those are improvements you can observe, which is the point.
Where this goes next
The threat side of this is getting harder, not easier. The NCSC's assessment of AI's near-term impact on the cyber threat expects more convincing social engineering at greater volume, which means the tells people were taught to look for are the first thing to disappear. Bad grammar was never a security control; it was just a lucky accident of attacker effort.
That is the case for measuring rather than assuming, and for measuring repeatedly. If you want the detail of how, human risk management covers the discipline, measuring security progress covers the numbers worth tracking, and the online scams guide covers what is actually landing on people right now.
Frequently asked questions
Why did CyberWatch AI get built?
Because people kept losing money to attacks that were obvious afterwards, and there was no fast, private way to ask whether a message was real. Inside organizations the same gap existed at scale: budgets went on technology while attacks arrived through the inbox.
Why do convincing scams work on careful people?
Modern scams are written to bypass judgment rather than defeat technology. The logos are right, the sender looks familiar, and the request is urgent enough that pausing to check feels rude or slow.
Why not just rely on the email filter?
Filters remove a large share of attacks, which is exactly why what reaches a person is the small fraction designed to survive filtering. The message that lands in the inbox has already passed the control you were relying on.
Why is a free scanner part of a business product?
The question people most need answered costs nothing to ask and does not scale as a paid feature. Checking a link, message or screenshot is free with no daily limit, and the organization console is the paid product.
Why measure people rather than just train them?
A training completion record shows attendance. It cannot tell you how a specific team will respond to a specific lure, which is the only thing that matters when the attack arrives.
Does CyberWatch AI use scanned content to train its models?
No. Scanned links, messages and screenshots are analyzed then discarded, only the risk score and a short summary are kept, and nothing scanned is used to train AI models.
Who is CyberWatch AI for?
The scanner is for anyone with a suspicious message. The console is for organizations that need to know how their people would respond, particularly those without a dedicated security team to find out.
Sources
- Annual Reports, FBI Internet Crime Complaint Center
- Internet Organised Crime Threat Assessment (IOCTA), Europol
- Telling users to 'avoid clicking bad links' still isn't working, UK National Cyber Security Centre
- Cyber security breaches survey 2025, UK Department for Science, Innovation and Technology
- Small organisations guide to cyber security, UK National Cyber Security Centre
- The near-term impact of AI on the cyber threat, UK National Cyber Security Centre
- ENISA Threat Landscape 2025, European Union Agency for Cybersecurity


