Samsung engineers put chip-making secrets into ChatGPT. No one hacked them.

There is no intrusion in this story. There is a text box.

What went into the prompt

According to reporting from Economist Korea and follow-ups in the trade press, the three incidents looked like ordinary engineering work:

1. An engineer pasted source code from a program that talks to semiconductor equipment measurement systems and asked ChatGPT to find the bug.
2. Another employee uploaded code used to spot defective equipment and asked for optimization.
3. A third pasted a confidential meeting record and asked for minutes.

Samsung had lifted an earlier ban on 11 March so staff could use the tool. After the leaks it imposed an emergency cap of 1,024 bytes per prompt and warned that another incident could mean a network-level block. A wider ban on generative tools followed.

Why this is a security incident

ChatGPT’s consumer terms at the time were not an NDA. Prompts could be reviewed and used to improve the system. Once the text left Samsung’s network, Samsung could not pull it back.

The data was not “stolen” in the usual sense. It was transmitted by people trying to go faster. That is the part other companies should copy onto a whiteboard. The threat model is not a nation-state in the source repo. The threat model is a competent employee with a deadline and a public model.

Chip process code and internal meeting notes are exactly the class of material competitors would like and regulators would call a trade secret. Volume does not have to be huge for the event to matter. One good paste is enough.

What to do that is not “ban everything forever”

A ban is a control. It is also a control employees route around on their phones. The more durable version is:

• Say which data classes may never go into an external model.
• Give people an internal tool if the company actually wants the productivity.
• Log and train as if this will happen again, because it will.

Call it shadow AI if you like. The mechanism is older than ChatGPT: staff used the convenient system, not the approved one.

This will not be the last time a headline about “AI security” is actually a headline about data leaving through the front door.