A principal at a 60-person engineering firm told me last month that one of his drafters pasted a client’s site plan into ChatGPT to ask for a takeoff shortcut. He found out because the drafter mentioned it in a standup, casual, like it was nothing. No incident report. The client never knew. When I asked what the firm’s policy was on pasting client drawings into a hosted model, the principal said they did not have one. They had a policy about not sharing files externally. The drafter did not connect a chatbot to “externally.”
That is the shape of the Samsung leak, just at a smaller balance sheet.
What happened at Samsung
In early 2023, engineers at Samsung’s semiconductor division accidentally leaked sensitive internal data by pasting proprietary source code and confidential meeting notes into ChatGPT, per reporting on the incident. The incident prompted Samsung to quickly implement restrictions or outright bans on generative AI tools.
Samsung had the budget to detect the leak, issue a memo in days, and roll out a restriction across a global workforce. Most engineering firms I work with would not catch it until a client asked why their drawing showed up somewhere it should not.
The 42% problem is not a Samsung problem
Metomic found that 42% of enterprise data leaks in 2024 were traced back to the use of public AI services with sensitive information. That is a measurement of leaks already observed, not a forecast. The Samsung incident is an early, well-publicized example of a pattern that has since become ordinary.
The engineering firms I talk to hold the same exposure shape as Samsung. Proprietary site plans. Client designs under NDA. Sequences and process data. Calc packages. Tooling geometry. All of it is exactly what a tired engineer will paste into a chat window if the sanctioned path is unclear or does not exist.
What is not on the table for a 60-person consultancy is Samsung’s incident-response budget. No security operations center. No legal team on retainer to write a breach notification letter at 11 p.m. Usually no enterprise agreement with a no-train clause in place when the leak happens. The cost of the same leak is higher for them, not lower.
Samsung had the budget to detect the leak, roll out a company-wide restriction, and survive the disruption. A 60-person engineering firm has no internal AI team and no private model. The incident-response capacity Samsung could throw at the problem is leverage the rest of us do not have.
A ban does not work the way you think
Most firms that catch a leak like this reach for the obvious lever: ban the tool. Samsung did it. It reads as decisive in a board meeting. It also fails inside an engineering team.
Engineers under deadline pressure do not stop using AI because a policy says so. They move to personal accounts, personal devices, and personal API keys. The same data still leaves the boundary. What you lose is the audit trail that would have told you it happened, and any chance of catching the second leak before a client does. I wrote about this in a companion piece on keeping IP and client data out of AI tools: a ban without a boundary relocates the risk to a place you cannot see it.
What works instead
The fix is two documents, not one ban.
First, a sanctioned-tool policy. Name the tools your firm has actually approved, at the tier each one is approved for, with the enterprise agreement or no-train clause in place where it matters. Make it short enough that a drafter can read it in two minutes.
Second, a data-class boundary. Sort your work into three buckets: public, internal but not client-identifying, and client-confidential or pre-patent. The client-confidential bucket never leaves a boundary you control. The middle bucket can go to an enterprise tier with a no-train clause. The public bucket can go to a hosted model with reasonable terms. Once those documents exist, the conversation with counsel stops being adversarial. You are asking them to review a boundary, which is the thing they are equipped to do.
The smaller-firm version of the same exposure
The principal I mentioned at the start called me a week later. His firm had pulled chatbot access for the drafting team and told everyone to stop. He wanted me to know the team was still using it, just on personal accounts, and he had no way to tell which drawings had gone where. That is the predictable outcome of a ban without a boundary. Samsung could absorb it. His firm cannot.
If your team is sitting between “we need AI to stay competitive” and “legal will not sign off,” that gap is usually a half-day of scoping work, not a six-month evaluation. The AI Workflow Diligence Sprint is a fixed-fee scoping call where we name your sanctioned tools, draw the data-class line, and hand you the two documents counsel can actually review.