AI-generated code exposed 95,364 email addresses
A missing bracket in an AI-written script showed 95,364 members' email addresses to other members. Singapore's regulator blamed the missing checks.
Published · 4 min read

On 25 April a Singapore company sent its members a marketing email in batches of 1,000, and each member's address could be seen by up to 999 others. An employee had written the Python script that sent it with a generative AI tool, and one missing bracket put each batch into a single "To" field. Singapore's Personal Data Protection Commission published the case on 21 September, and in the week of 28 September it reached the press: 95,364 members of Bee Cheng Hiang, known for its bak kwa, had their email addresses exposed. The useful part is the regulator's finding: the AI tool did not malfunction.
What failed was everything around the tool. The prompt asked for a script to send "mass email using a local list" and never said that members must not see one another. The employee tested it by checking the activity logs rather than opening the test email, so the error went unnoticed until the real send. Nobody else reviewed the work, and the company had no policy on staff using generative AI; this was its first attempt to use AI tools in its operations. Mothership reported it as the first AI-related breach the PDPC had been notified of.
The fixes are checks, not a better tool
The company stopped the mailing, fixed the script and told every affected member, and at least two staff now check each bulk email before it goes out. Under the undertaking it signed on 2 September it is going further: independent technical review of any AI-generated code that involves personal data, test sends to dummy accounts first, an automatic block on emails carrying more than one recipient in a single entry, a written breach procedure, and training for the staff who build and review its systems. None of that changes the AI tool. Three of those measures would have caught this mistake before any member saw it.
That is the lesson for an office here. Anyone can now ask an AI tool for a script, a spreadsheet formula or a mail merge and get something that runs. Running is not the test. The script did what the prompt said, and the prompt left out the rule everyone assumed.
Test it the way a member would see it
Every office keeps lists of people: members, staff, customers, applicants. Code that sends anything to one of those lists is the code I would least want to run untested, whoever or whatever wrote it. The test that would have caught this one is cheap. Send to a handful of addresses you control, open what arrives, and read the To line the way a member would.
I hold the AI in my own company to the same rule. Before Storia's assistant was allowed to change a couple's guest list, I ran 150 scripted conversations against it, 27 of them deliberate attempts to trick it, and none got a change saved without the couple confirming it. The first round still found faults: it could be pushed into claiming it had deleted guests, and into repeating its own instructions. Both were fixed, and both held in the rounds after. It is the principle I wrote up in the AI proposes, a person confirms: the AI does the typing, and a person keeps the decision.
Five checks before AI-written code touches real records
- A second person reads it. Any script, formula or macro that touches personal data is reviewed before it runs by someone who did not write it, whoever wrote the first draft.
- Test on records that are not real. Send to a few dummy accounts your office controls, then open the result as a recipient would. A log that says 1,000 sent says nothing about who saw whom.
- Write down what must never happen, not only what should. "No member sees another member's address" belongs in the prompt and in the test.
- Prefer the tool built for the job. A newsletter service is built to send each member a separate copy; a script written in an afternoon has to be told to.
- Write one page on AI at work: which tools staff may use, what data may go into them, and who reviews what comes out. The company in Singapore had none.
If nobody in your office would have caught the missing bracket, that is worth knowing before the next mailing, and the fit check is a short way to start.
Sources
- Personal Data Protection Commission (Singapore), Voluntary Undertaking by Bee Cheng Hiang Marketing Pte Ltd, 21 September 2026
- Mothership, 95,364 Bee Cheng Hiang members' data exposed after employee used AI to send mass email, 1 October 2026
- AsiaOne, Nearly 100,000 affected as Bee Cheng Hiang suffers Singapore's first AI-related data breach, 1 October 2026
Enrique Lacambra III builds internal business systems from Cainta, Rizal.
Tell me what is broken and I will tell you if it is worth building. A 20-minute call, no charge, and I will say so if you do not need me.
