"SMEs Have No Data Worth Stealing" — The Most Dangerous Myth in AI Adoption

In September 2026, the PDPC announced Singapore's first AI-linked data breach. No hacker was involved. At a household-name local food brand, an employee asked a generative AI tool to write a mass-mailing script; the prompt never said to hide recipients from each other, the generated code skipped BCC, and 95,364 customers' email addresses were exposed — one employee, one AI tool, one missing instruction, zero review. Remember that story every time you hear the line we keep hearing from AI startup founders: "SMEs don't have data anyone wants, so there's no real risk in putting everything through a cloud LLM." It's wrong on both halves: your data is more valuable than you think, and "someone wanting it" was never the main risk in the first place.

To be clear about where we stand: we run an AI-driven MSP and we push clients toward AI adoption, not away from it. This isn't an anti-AI post. It's a correction of the reasoning — because a good decision made for a wrong reason eventually becomes a bad decision.

Where the myth comes from

The logic goes: we're not a bank, we don't hold state secrets, our "secret sauce" wouldn't interest anyone. Therefore nobody is going to steal our prompts from OpenAI's servers, so why pay for enterprise plans or worry about data controls?

The mistake is picturing a human attacker deciding whether your company is "worth it". That's not how any of this works. Modern attacks are automated and indiscriminate — nobody needs to want your data specifically, the same way a phishing botnet doesn't care whose mailbox it lands in. And as Singapore's first AI breach shows, the biggest AI data risks don't involve an attacker at all: the data walked out through an unreviewed script, not a break-in.

What your "boring" data is actually worth

The breached company was a food brand — bak kwa and gift boxes, about as far from a bank as business gets. It still held nearly 100,000 customer email addresses, and losing them made national news. Take a typical 30-person Singapore SME and list what actually sits in its systems — and now, in its staff's chat histories with AI tools:

• Customer PII — names, NRIC numbers, phone numbers, addresses. Personal data appears in roughly 46% of all breaches because it sells, powers scams, and fuels identity fraud. Your customers' data is exactly as valuable as a bank's customers' data — it's the same people.
• Credentials — still the single most common way attackers get in anywhere. Staff paste config files, connection strings and API keys into chatbots more often than anyone admits.
• Your clients' secrets — the quotation you're preparing for a listed company, the payroll file you process for a client, the M&A discussion in your mailbox. SMEs are the documented backdoor into bigger organisations: attackers compromise the small vendor to reach the large customer. You may not be the target — you're the door.
• Your own money and name — the average breach costs a Singapore SME around S$120,000 in local industry estimates, and the reputational bill arrives separately. Ask any brand that has had to explain itself in the newspapers.

What actually happens to data you give a cloud AI tool

"Being stolen by hackers" is the least likely fate. These five are the realistic ones:

A useful gut-check before any paste: would you email this exact content to a stranger who promises to keep it safe? Because mechanically, that's what pasting into an unmanaged AI tool is.

1. On consumer plans, your chats train the model — by default

As of 2026, the free and personal-paid tiers of every major AI assistant — ChatGPT, Gemini, Claude, Copilot, Grok — use your conversations for model training by default. ChatGPT's "Improve the model for everyone" toggle ships switched on; Anthropic moved consumer accounts to default-on training in September 2025. Business, Enterprise and API tiers are the opposite: no training by default. That single tier difference — not the model, not the price — is the main data-protection line. Most SME staff are on the wrong side of it, on personal accounts the company can't see.

2. Your prompts sit in retention on someone else's infrastructure

Whatever you paste is stored, for weeks to indefinitely depending on plan and settings. Stored data inherits every risk of the provider: their breaches, their misconfigurations, their legal obligations. AI providers have already had chat histories exposed by database misconfigurations, and court orders have forced providers to preserve user conversations they would otherwise have deleted. Your data's safety is now someone else's operational discipline.

3. Chat history turns one stolen password into a full archive

A year of an employee's AI chat history is a searchable archive of everything they worked on: client names, pricing, contracts, internal disputes. One phished password on a personal ChatGPT account — usually without MFA — hands all of it over at once. This is the "stolen data" scenario that actually happens, and it doesn't require anyone to target you.

4. The AI's output can leak data all by itself

This is the one Singapore's first case made real. The risk wasn't what the employee typed in — it was what the AI produced: a working script with one safeguard missing, run against a live customer list with nobody checking it first. AI output is confident, fluent and occasionally wrong in ways a non-specialist won't spot. Code, mail-merges and automations that touch customer data need a human review before they run — the breach was one code review away from never happening.

5. The AI wrapper you bought has its own data problem

Here's the uncomfortable part for the founders repeating this myth: many SME-facing AI products are thin layers over the same cloud LLMs, and the data handling behind them is often undocumented — which API tier, what logging, what retention, which country. When a vendor answers your data question with "SMEs don't have data anyone wants", that's not a security architecture. Ask instead: where does my data go, who stores it, for how long, and is it used for training? A serious vendor answers in one paragraph.

The PDPA point nobody mentions

Even if you accept the myth entirely — suppose nobody on earth wants your data — the compliance obligation doesn't care. Under the PDPA, pasting customer personal data into a cloud AI tool is a disclosure and, for overseas-hosted services, a cross-border transfer. Both need to satisfy the Act regardless of whether any attacker is interested, and the PDPA applies to a 5-person company the same as a 5,000-person one — with penalties up to 10% of annual Singapore turnover or S$1 million. In the September case, no hacker ever appeared, yet the company still ended up with a PDPC undertaking and its name in the headlines — the obligation bit anyway. The PDPC's guidance on AI puts responsibility on the organisation deploying it — you — not the model provider. We covered the fundamentals in our PDPA guide for Singapore SMEs.

The right conclusion is not "avoid AI"

The opposite overcorrection — banning AI tools — fails too, and predictably: staff just use personal accounts on their phones, and you lose the productivity and the visibility. Surveys of workplace AI use find most sensitive pastes already happen through personal accounts outside company oversight. Note that the September breach wasn't caused by using AI — it was caused by using AI with no rules around it. The goal isn't less AI. It's the same AI, through accounts and tiers you control, with someone owning the rules.

A minimal AI usage policy for an SME — five rules

You don't need a 40-page governance framework. You need these five rules, written down and enforced:

1. Company AI accounts only. Business-tier ChatGPT / Claude / Copilot, provisioned and off-boarded like email. No work data in personal accounts — this one rule eliminates the training-by-default problem and most of the visibility problem.
2. Classify before you paste. A one-line rule staff can remember: customer personal data, credentials, and client-confidential documents don't go into any AI tool without approval. Everything else is fine.
3. Review AI output before it touches customers. Any AI-generated code, mail-merge or automation that handles customer data gets a second pair of eyes before it runs. This is the rule that would have prevented Singapore's first AI breach — one review, 95,364 addresses safe.
4. MFA on every AI account, and vet AI vendors with four questions. Chat history is an archive of your business — protect it like your mailbox. And before buying any AI product: where is my data processed, who stores it and for how long, is it used for training, can you delete it on request? No clear answers, no deal.
5. Name an owner. Someone — internal or external — owns AI usage: which tools are approved, who has access, and reviewing it quarterly. Governance that belongs to nobody doesn't exist — and "nobody owned it" is the one-line post-mortem of most AI incidents, including September's.

Frequently asked questions

Is it safe for my business to use ChatGPT?

Yes — on a business or enterprise tier, with MFA, under a simple usage policy. The risk isn't the tool; it's free personal accounts handling customer data with default settings, and AI output running unreviewed.

Does ChatGPT or Claude train on what my staff type?

On free and personal paid plans — yes, by default in 2026, unless the user opts out. On business, enterprise and API tiers — no, not by default. Check the tier, not the brand.

What actually happened in Singapore's first AI data breach?

An employee at a local food brand used a generative AI tool to write a bulk-email script. The prompt didn't specify hiding recipients, the code skipped BCC, and 95,364 customer email addresses were exposed to other recipients. The PDPC announced it in September 2026 and accepted the company's voluntary undertaking to improve PDPA compliance. No attacker, no theft — an unreviewed AI output and a missing rule.

Does the PDPA really apply to a company our size?

Yes. The PDPA has no small-company exemption. If you hold personal data of customers, staff or prospects, the obligations — including how that data flows into and out of AI tools — are yours.

An AI vendor told us our data isn't sensitive enough to worry about. Red flag?

Yes. It means they either haven't thought about data handling or hope you won't ask. Put the four vendor questions above to them — the answers (or the silence) tell you everything.

Our approach

We sell AI adoption for a living — which is exactly why we won't repeat the "nobody wants your data" line to win a deal. Evernet's VCAIO service exists to give SMEs the adoption and the governance in one motion: the right tools on the right tiers, a usage policy people actually follow, output review where it matters, and PDPA obligations handled from day one.

Rolling out AI and not sure where your data is going? Book a discovery call — we'll map where AI is already being used in your company (there's always more than you think) and what to fix first.