AI Bubble or Not, Security Is the Part That Won't Pop

Nobody knows if AI is a bubble. Smart people disagree, and the numbers point both ways. But one part of the AI story will not pop: security. Businesses that rush AI tools in are opening new holes right now, and those holes stay open whether the market goes up or down. The businesses that build safely are the ones still standing either way.

That is the short answer. Here is the longer one.

The bubble question, both sides

Business owners ask me this almost every week. Here is what each side points to, from recent sources. I am not going to tell you who is right on the economics. I build automations, I don't predict markets.

The case for "bubble":

Spending far ahead of revenue. In October 2026, Gartner said AI data center owners need about $8.5 trillion in revenue through 2030 to cover their spending, their financing and a 12% return. It expects at most $3.6 trillion of AI revenue to reach those data centers (Fortune, 17 Oct 2026).
A classic boom pattern. Bank of America said the AI boom follows a capital cycle that usually ends in disappointment, "even though the underlying technology often lived up to expectations" (same Fortune report).
Debt and circular deals. The same report points to heavy borrowing, rising interest rates, stress at one large AI cloud provider, and "circular" deals where AI companies fund each other and then buy from each other.

The case against:

Real sales, not just hype. Taiwan's TSMC, a major maker of AI chips, reported record revenue of NT$1.49 trillion for July to September 2026. Analysts put it down to strong AI demand (Focus Taiwan, 8 Oct 2026).
Earnings keep growing. Fortune notes that earnings kept rising and stocks came back to record highs after earlier bubble warnings. One strategist asked, "How many canaries have we killed in the coal mine over the last three years?"

So there is a real argument on both sides. Nobody can promise you which way it goes.

Why it doesn't matter much for your business

If you run a small business, you are not betting billions on data centers. You are using AI tools: a chat bot on your website, an assistant that writes emails, an automation that moves leads into your CRM.

Those tools work the same whether AI stocks go up or down. What changes your week is something else. Is your setup safe? A leaked key or a hijacked email account costs you money and trust today. That risk is already real, and it is growing.

The new holes AI brings

These are the ones I see most. Each one is backed by a current source.

1. API keys leaked in code and files. An API key is a password that lets one tool use another. GitGuardian found 28.65 million new secrets pushed to public GitHub in 2025, up 34%. Leaks of AI service keys went up 81% (GitGuardian, Mar 2026). Over 64% of keys found in 2022 still worked when tested again in January 2026. People leak keys and never cancel them.

2. AI bots and agents with more access than they need. OWASP, a respected security group, calls this "excessive agency" and lists it in its top 10 risks for AI apps (OWASP LLM06). IBM found that about 1 in 5 breached organizations had an AI-related breach, and 92% of those lacked proper access controls on their AI (IBM, Jul 2026).

3. Prompt injection. This is when a bot reads an email or web page with hidden instructions and obeys them. OWASP lists it as risk number one (OWASP LLM01). The UK's National Cyber Security Centre warned it may never be fully fixed, so the goal is to limit the damage (NCSC, Dec 2025). A bot that can only draft a reply can't do much harm. A bot that can send, delete and pay can.

4. Staff pasting client data into AI tools. It is fast and it feels harmless. But the rules differ by plan. One major AI company's consumer plans can keep chats for up to five years if the training setting is on. Its business and API plans are not covered by that change (Anthropic, Aug 2025). If staff use personal free accounts, you don't control where client data ends up.

5. Fake AI apps and AI-written phishing. Sophos reviewed a year of real cases and found criminals posing as AI products in 30 of 38 AI-related incidents, with fake apps and browser extensions (Help Net Security, Aug 2026). Microsoft found AI-written phishing emails got a 54% click rate, against 12% for normal ones (The Register, Oct 2025). The old advice to look for bad spelling no longer works.

6. AI vendors' data policies. Every AI tool you connect has its own rules on what it keeps, for how long, and whether it learns from your data. Those rules change, as the example in point 4 shows. Someone in your business should read them before client data goes in.

Lessons from my own work

I build automations for businesses every day. These are things I found and fixed. No client names here, only the lessons.

Keys in plain files. On more than one project I found API keys saved in plain text inside project folders and notes. I moved every one into an encrypted vault on my PC. Now a script gets a key only while it runs, and the key never sits in a file.
Only the access a key needs. When an email service offers a "sending only" key, that is the one to use. If that key leaks, someone can send email, but they can't read your contacts or change your account.
2-step login that phishing can't beat. A code by text message is better than nothing. A passkey or security key is better still, because it won't work on a fake login page. The US cyber agency CISA calls this phishing-resistant MFA (CISA fact sheet).
Cloud servers block email ports by default. I hit this setting up my own automation server. DigitalOcean blocks email ports 25, 465 and 587 on new servers (DigitalOcean docs). That is a good default. The safe fix is a proper email service, not opening the ports.
Backups that tell you when they fail. A backup that fails quietly is worse than none, because you think you're covered. My server backup reports in every night. Every morning a check looks for that report, and if it's missing or failed, I get an email.
Never run a stranger's "test project". Fake recruiters send developers a "coding test" that installs malware. Voice of America reported one group hit at least 30,000 computers in over 100 countries this way (VOA, Oct 2026). I never run unknown code on the PC that holds client work. If I have to look at it, it goes on a separate machine.

Your 10-point checklist for this month

You can do most of this yourself in a few afternoons.

List every AI tool your business uses, including the free ones staff signed up for.
Find your keys. Search your shared drives and notes for API keys and passwords. Move them into a password manager or vault.
Cancel keys you don't use, and replace any key that was ever pasted in a chat or email.
Cut access down. For each AI tool or bot, ask "does it need to send, delete or pay?" If not, give it read-only or draft-only access.
Keep a human in the loop for anything that sends money, deletes data or emails all your customers.
Turn on 2-step login everywhere, and use passkeys or security keys for email, your CRM, your domain and your bank.
Write one short rule for staff: what client data never goes into an AI tool, and which approved business account to use instead.
Get AI apps only from the official site or app store, and check the publisher name before you install anything.
Read the data terms of every AI vendor you use. Turn off training on your data where you can.
Test your backups and make sure a failed backup sends an alert to a real person.

Bubble or not

The market will do what it does. Your setup is the part you control. Build it safely, and you're fine whichever way the bubble question goes.

Want me to check your setup? Book a free automation audit. I'll go through your tools, keys, logins and backups with you, and you get a short list of what to fix first. You can also see what my website security work includes.

Questions

Is AI a bubble?

Nobody knows yet. Some analysts point to huge spending, debt and circular deals between AI companies. Others point to record chip sales and strong earnings. For a small business owner the more useful question is whether your own AI setup is safe, because that risk is real either way.

What is the biggest AI security risk for a small business?

Usually not a movie-style hack. It is the plain stuff: API keys sitting in files, AI tools connected with full access to email or the CRM, staff pasting client data into free AI apps, and phishing emails that now read perfectly.

What is prompt injection, in plain words?

It is when an AI bot reads an email, a web page or a file that has instructions hidden in it, and follows them as if you gave them. OWASP lists it as the number one risk for AI apps. The fix is to limit what the bot is allowed to do, so a bad instruction cannot do much harm.

Do I need to stop using AI tools?

No. Use them, but set them up like you would a new staff member. Give each tool only the access it needs, use business plans with clear data terms, keep keys in a vault, and turn on a 2-step login that a fake page cannot steal.

How long does a security check of my automations take?

For most small businesses, the first look takes one call. I go through your tools, keys, logins and backups, and you get a short list of what to fix first. Book a free automation audit to start.

Related service

Website Security for Small Businesses: Hardening, hacked-site cleanup and incident response for small business websites: Cloudflare rules, WordPress login protection, keys and email security.