A couple of brackets

Singapore’s first reported AI-related data breach came down to a couple of brackets.

What happened

In April, an employee at Bee Cheng Hiang asked an AI tool to write code that would send a marketing e-mail to the customer list in batches. The prompt didn’t say to hide each recipient’s address. So every e-mail carried 1,000 customers’ addresses, and more than 95,000 people’s addresses were exposed.

The AI didn’t malfunction. PDPC called it “human error in developing the e-mail distribution code with an AI tool.” The code was tested by checking activity logs, not by reading a test e-mail. One person wrote it, with no review and no policy on using AI for work.

The bucket: vibe-coded, and not meticulous

It’s both. The code was written quickly with AI and nobody checked it. And 95,000 addresses sat in a list that a single untested script could send to everyone.

Where AI comes in

Most of this newsletter is about attackers using AI to find old weaknesses cheaply. This one is the other side: AI made it cheap to make the weakness. Writing code that touches customer data now takes minutes. Reviewing it still takes a person who knows what to look for, and that’s the step that was skipped here.

What I’d do

The fix PDPC accepted is a good template:

• An independent technical review of AI-generated code that touches personal data.
• Two people checking every bulk e-mail.
• Test sends to dummy accounts first - and someone reads the test e-mail, not just the logs.
• An automatic block on any e-mail carrying many addresses, so it doesn’t depend on anyone remembering.

And if you’re a customer: expect messages that look like they come from brands you deal with. Check before you click.