Hello, this is Ryuta Hamamoto from TIMEWELL.
Building a business tool with AI is no longer unusual. Tell a tool like Claude Code "I need this kind of report" and you get something that runs in a few minutes. One pattern is spreading fast: pack every feature into a single HTML file and pass it around over Slack or email.
No server, no deployment, and you throw it away when you are done. For the person making it, nothing is lighter. I have seen posts on social media saying that once you get used to this, the very idea of deploying disappears. I understand the feeling. I have been saved more than once by a quick throwaway tool.
From the security and IT side, though, the same scene looks very different. Building has become so easy that governance has not kept up. This column breaks down five reasons it is risky, compares it with the Excel macro era, and sets out how an organization can respond.
Convenience Is Creating Apps Nobody Can Count
In IPA's "10 Major Security Threats 2026" for organizations, "cyber risks surrounding the use of AI" entered at third place for the first time.1 IPA picks candidates from the previous year's incidents, and about 250 researchers and corporate practitioners vote on them. The ranking tells me that people in the field now feel the scope of what defenders must watch has widened.
Hearing "AI risk," most people think of confidential data typed into a chatbot. I think an equally large problem is what AI produces circulating inside the company on its own. I covered shadow AI in a separate article, the reality of shadow AI and how to design an internal policy. This one looks at the next step: apps that AI builds.
What makes a single HTML file tricky is that it skips every procedure normal software goes through. No request, no development process, no approval to publish. Anyone can build and distribute one, anytime, for free.
Five Reasons It Is Risky
1. Apps nobody has counted keep multiplying
How many HTML files exist, and whose hands are they in? The IT department has no way to know. There is no register and no detection mechanism.
Protection starts with counting what you protect. Something you do not know exists never gets patched, inventoried, or retired. Each tool may be small, but you can easily end up with hundreds of ownerless apps drifting around the company.
2. API keys and internal URLs get embedded in the file
When you try to finish everything in one file, it is tempting to write API keys, access tokens, and internal endpoints straight into the code. For getting something to run, that is the quickest way.
But an HTML file is opened in a browser, so anyone can read its contents. The moment you hand the file over, you hand over the secrets with it. MITRE's weakness catalog lists this as CWE-798, use of hard-coded credentials, and notes that when the credential sits on the client side, extracting it is usually very simple and the threat is greater.2
Later, you cannot fix it cleanly. You do not know how far the file was forwarded, so you cannot tell which keys need to be revoked.
3. Flaws are found, but cannot be fixed
Distributed files scatter across personal PCs, mailboxes, and chat histories. When a vulnerability turns up, there is no way to collect every copy and replace it.
"Throw it away when you are done" is a fair instinct for the creator. In practice, though, you only throw away your own copy. Old versions stay in someone's Downloads folder indefinitely.
4. You cannot trace who used it or what data it touched
Unlike something hosted on a server, an HTML file has no usage records and no access control. Who can use it is decided by how far it was forwarded.
When a leak happens, the first questions are who accessed what data, and when. Without records, you cannot identify the cause, explain the impact, or report to customers and partners.
5. Sloppy AI-written code spreads without anyone checking it
AI-generated code can contain vulnerabilities. In a study Veracode published in July 2025, more than 100 large language models were given 80 coding tasks, and in 45% of cases the code introduced a vulnerability in the OWASP Top 10 categories. The models failed to defend against cross-site scripting (XSS, where a malicious script runs on someone else's screen) in 86% of the relevant cases.3 Failure rates for Python, C#, and JavaScript fell between 38% and 45%.
A caution on reading the numbers: this is one set of tasks, and it does not directly measure how risky the tools built in real workplaces are. Still, it shows clearly that writing code that works and writing code that is safe are different skills.
The organizational problem is that the creator often has not read the code, and nobody else reviews it before it spreads. "It runs" and "it is safe" are different things.
AI Security training, taken seriously
A 2-day intensive course fully aligned with OWASP, NIST, ISO/IEC 42001, and METI. Take it as executives, practitioners, or both.
The Excel Macro Problem, but Faster
This pattern is not new. Someone on the floor builds an Excel macro or an Access database, and before long the business depends on it. When that person moves to another department, nobody understands what is inside. Many companies have lived this already.
AI-built HTML apps are an extension of that story, but the impact grows in three ways.
| Aspect | Excel macros and Access | AI-built HTML apps |
|---|---|---|
| Who can build | A few people who can write macros | Everyone who can write instructions |
| Build time | Days to weeks | Minutes to hours |
| What it connects to | Mostly local files | External APIs, cloud, internal systems |
The third row matters most. A macro's effects mostly stayed inside its file. An HTML app connects directly to outside services and internal data, so when something goes wrong, the damage can reach beyond the company.
There is one more point. As the number of people who can build grows, the share of people who can tell something is dangerous shrinks. Someone who wrote macros at least knew they were programming. Today, programs are being handed around by people who do not realize they made one.
The Response: Build a Safe Path Rather Than Banning
I do not recommend a blanket ban. The productivity gain is real, and a ban only pushes use out of sight, which is more dangerous. What matters is creating a state where doing it the official way is easier.
1. Provide an official place to publish
Instead of passing HTML around as files, put it in an internal location and share a URL. Require sign-in with a company account, and record the creator, purpose, and data handled in a register at the moment it is published.
That alone resolves much of reasons 1, 3, and 4. Replacing the page means everyone uses the new version, usage is logged, and you can count who has published what. It does not take anything exotic: a place that can host internal static pages, plus sign-in.
2. Keep secrets out of files
When an external API is needed, set up a relay server and keep the keys there. The HTML holds no keys. In addition, scan automatically before publishing for keys and passwords.
Building this rule into the instructions you give AI and into internal templates helps a lot. If the opening prompt already says "never write API keys in the code" and "pin the version of any external script," the rule is more likely to be followed without the creator thinking about it.
3. Draw the line by the data involved
If you review every tool with the same rigor, people stop using the official route. Vary the weight of review by how sensitive the data is.
| Data handled | Examples | Required procedure |
|---|---|---|
| Public information only | Calculators, internal study materials | Build freely. Just register on the official site |
| Internal information | Business summaries, case management | Register and pass the automated checks |
| Personal or confidential data | Customer data, HR records, partner secrets | Mandatory review by the responsible department |
The key to this table is keeping the top row as light as possible. Make it heavy and even the harmless tools go underground. Only the bottom row should be heavy.
4. Use AI for the review
Now that building is faster, human review alone cannot keep up. Add a step before publishing where AI checks for vulnerabilities and embedded secrets. People then focus on what the AI flags and on high-importance tools.
Some people are uncomfortable with AI checking what AI built. I find it realistic to treat this as a first-pass screen. Humans keep final responsibility, and the AI frees up their time for the places that are truly dangerous.
5. Set an end date and take inventory regularly
Set an expiry when a tool is registered, and when it arrives, ask the creator whether to continue or retire it. A tool that stays in use and gains users is a candidate to be rebuilt as a proper business system.
A tool built on the floor is a source of risk, and also a prototype that shows which functions are truly needed. Whether you can pick up on that decides if the relationship between IT and the business shifts from watcher and watched to building together.
What You Can Start on Monday
Start with these three. None needs a large budget.
First, learn the actual situation. Have the IT department look at how many HTML files are moving through Slack and email. It is usually more than you expect. Second, choose a candidate for the official publishing place. An existing internal portal or file-sharing service is enough to start. Third, add one line to your AI instruction template: do not write secrets in the code.
For a rough read on where your company stands, the AI literacy check takes about three minutes.
Closing: The Problem Is Not the Format, but Having No Place and No Ending
The problem is not HTML itself. Building became easy, but where these tools live, how secrets are handled, and how they are retired were never decided.
Employees improving their own work is a real strength for an organization. To use it, I believe the answer is not prohibition but making the safe path the easiest one. I would like to apply the lessons from the Excel macro era ahead of time, this time.
How many HTML files are circulating in your company right now? Try counting them once.
How to Respond to This Change
Letting the floor build its own tools is something to shape, not stop. Where to publish, how to treat secrets, where review begins by data sensitivity, how to run inventories: you do not have to decide everything at once. Setting priorities and fixing things as you go is usually faster.
TIMEWELL's WARP SECURITY supports the design and operation of security around AI use. How to take inventory of floor-built tools, where to leave freedom and where to start review, how to build AI into the review itself: we think it through with you, starting from priorities for your situation. For the builder's side of AI use, see also our practical guide to security for vibe coding on Vercel.
If you want to first learn what is actually circulating, or to set lines that work without banning, feel free to reach out.
Footnotes
-
IPA, "Information Security 10 Major Threats 2026" announcement (January 29, 2026, in Japanese). The first-time entry of "cyber risks surrounding the use of AI" at third place for organizations, and the selection method (IPA picks candidates from the prior year's incidents; about 250 members of the selection committee vote), are from this announcement. ↩
-
CWE-798: Use of Hard-coded Credentials (MITRE). The definition (a product containing hard-coded credentials such as a password or cryptographic key) and the statement that client-side systems with hard-coded passwords pose even more of a threat, because extraction is usually very simple, are from this page. ↩
-
AI-Generated Code Poses Major Security Risks in Nearly Half of All Development Tasks, Veracode Research Reveals (Veracode, July 30, 2025). The 80 coding tasks, more than 100 models, the 45% rate of introduced vulnerabilities, the 86% failure against XSS (CWE-80), and the 38 to 45% failure rates for Python, C#, and JavaScript are from this release, which describes the 45% as the share of cases introducing vulnerabilities classified in the OWASP Top 10. ↩






