In this guide · 7 sections
Short answer
An employee used a generative AI tool to write a bulk-email script that put up to 1,000 customer addresses in each email's To field. 95,364 people were affected. The PDPC accepted a voluntary undertaking and said the cause was the prompt and missing review, not the AI tool.
What happened in the Bee Cheng Hiang data breach
The facts below come from the PDPC's voluntary undertaking page for Bee Cheng Hiang Marketing Pte Ltd, published on 21 September 2026, and CNA's report of 30 September 2026, in which the PDPC called it the first AI-related data breach in Singapore that it had been notified of.
| Date | What happened |
|---|---|
| 25 April 2026 | A marketing email went to members in batches of 1,000, with every address in the batch shown in the To field |
| 27 April 2026 | The company notified the PDPC |
| 2 September 2026 | The voluntary undertaking was signed |
| 21 September 2026 | The PDPC published the case |
In total, 95,364 members were affected. Only email addresses were exposed, and the PDPC said there was no evidence of further misuse. The company stopped the email run, fixed the script, told every affected member and now requires at least two staff to check every bulk email before it is sent.
Why the AI-generated script leaked email addresses
An employee used a generative AI tool, on a personal account, to write a Python script that sent a "mass email using a local list" in batches through an email-sending service. The prompt did not say that recipients must not see each other's addresses. The script put all 1,000 addresses of a batch into one recipient group instead of one email per person. The PDPC page shows the code: the difference was a single misplaced bracket.
The employee tested the script by checking the sending logs, which showed emails going out successfully. Nobody opened an actual test email, which would have shown a To field with 1,000 addresses in it.
That framing matters for every business. Under the PDPA, your organisation is accountable for personal data in its care, whoever or whatever wrote the code, the formula or the email merge.
What the company agreed to do
A voluntary undertaking is the PDPC accepting a written commitment to fix the problems instead of going through a full investigation and decision. The PDPC verifies that the commitments are carried out. Bee Cheng Hiang's undertaking includes:
- A governance framework for AI-assisted coding, with independent technical review of AI-generated code that touches personal data, drawing on the AI Verify testing framework's principles.
- A secure development process, including test sends to dummy recipient accounts before any real email run.
- A written data breach response procedure with roles, escalation timelines and notification steps.
- Automated checks in the email system that block any bulk email with several recipient addresses in one message.
- Awareness training for all technical staff who build or deploy systems that handle personal data.
None of these need a large IT team. A small business can apply the same ideas in a few hours, scaled to its size.
Five controls for AI-written code, formulas and emails
You do not need to ban AI tools. You need the same checks you would apply to a new junior staff member's work. These steps follow directly from the PDPC's findings.
Write a one-page AI use policy
List which AI tools staff may use, on which accounts (company, not personal), and what must never be pasted in, such as customer lists or NRIC numbers. Make clear that any AI output that touches customer data is reviewed before use.
Have a second person review AI-written scripts
Any script, macro, Apps Script or automation that reads a customer list is checked by someone other than the person who prompted it. They read the code or flow, not only the result. If nobody in-house can read code, do not run it on real customer data.
Test with real test emails, not logs
Send the first run to two or three internal or dummy addresses and open the emails. Check the To, Cc and Bcc fields, the merge fields and the unsubscribe link. Logs only tell you that something was sent, not what the recipient saw.
Send bulk email through a proper marketing tool
Tools such as Mailchimp, Brevo or the email feature of your CRM send one message per recipient by design. Home-made scripts and Outlook mail merges are where the To-field mistake happens.
Require two-person sign-off before sending
For any email going to more than, say, 50 customers, one person prepares it and another approves it. This is the control Bee Cheng Hiang added straight after the incident.
If your business has a similar email slip
Exposed email addresses still count as a data breach. Stop the send, record what happened and when, and assess whether it must be reported. Under the PDPA you must notify the PDPC if the breach is likely to cause significant harm or affects 500 or more people, within 3 calendar days of deciding it is notifiable. Our guide to PDPA data breach notification walks through the assessment and the forms.
Email exposure: first steps
This is not legal advice. If the breach is large or involves sensitive data, speak to a lawyer and read the PDPC's own guidance.
Do it yourself or get help?
If you already have a data protection policy and a breach plan, add an AI section to them: approved tools, banned data, review rules and the bulk-email sign-off. Our PDPA checklist for small businesses shows where these fit, and our guide to spotting phishing emails covers the other side of staff training.
If you have none of this written down, our PDPA Compliance Pack (S$690) produces your privacy policy, staff data protection policy, data inventory and breach response plan from a short online questionnaire. Security Awareness Training (S$390) adds short lessons for staff, which can include safe AI use. We are not a law firm, and you remain responsible for your own compliance decisions.
Frequently asked questions
No. The PDPC said it was not a malfunction in the AI tool. The employee asked for a bulk-email script without saying recipients must be hidden, then tested by checking logs instead of opening a test email. The PDPC pointed to weak testing, no supervisory review and no policy on using generative AI at work.
The PDPC page describes a voluntary undertaking, not a financial penalty. The company committed to an AI coding governance framework, test sends before bulk emails, a documented breach procedure, automated email checks and staff training. The PDPC will verify compliance and can issue directions if the company does not follow through.
Yes. An email address that identifies a person is personal data under the PDPA, and showing it to other customers without consent is unauthorised disclosure. Whether you must notify the PDPC depends on the likely harm and on whether 500 or more people are affected. In this case 95,364 people were affected.
They can, but set rules first. Use company accounts, never paste customer lists or identity numbers into the tool, and have a second person review any AI-written code or automation that touches personal data before it runs. Test with dummy recipients and read the output yourself rather than trusting logs or the AI's explanation.
Stop sending bulk email from scripts or ordinary mailboxes. Use a proper email marketing tool that sends one message per recipient, send a test to internal addresses first, and require a second person to approve every send to customers. Those three steps prevent most To-field and wrong-list mistakes.