Samsung Engineers Leaked Secret Source Code into ChatGPT — Three Times in 20 Days
Summary
Within about 20 days of Samsung's semiconductor division allowing ChatGPT, engineers pasted confidential material into it three separate times: proprietary equipment source code (to debug it), defect-detection/yield code (to optimize it), and a recorded internal meeting (to generate minutes). Under OpenAI's terms at the time, submitted content could be used to improve models. The data could not be recalled. Samsung banned generative AI tools company-wide within a month.
Severity
SEV-1 — Proprietary semiconductor IP left the company's control irreversibly and entered a third party's systems where, under the terms in effect, it could be retained and used for training. Irreversible exposure of core trade secrets is maximum severity.
What happened
- Mar 2023: Samsung Semiconductor lifts its internal ban and allows engineers to use ChatGPT.
- An engineer pastes proprietary equipment database source code in to fix errors.
- A second engineer submits defect-detection / yield code to optimize it.
- A third feeds a confidential internal meeting transcript in for summarization.
- Samsung discovers three leaks within ~20 days; the data has already left its control.
- Within a month, Samsung bans generative AI tools on company devices and networks.
Root cause
Primary: Employees were given access to an external model that could ingest arbitrary data, with no technical control (only policy) preventing confidential IP from being sent outside the company boundary.
Contributing factors:
- The chat interface feels private, so well-meaning engineers didn't perceive that inputs left the company.
- Existing data-loss-prevention systems weren't watching the AI-tool channel.
- Access was granted before any interception or classification control was in place — controls came after the leak.
Governance analysis
The control that would have prevented it
A data-egress interception layer that classifies and blocks confidential content (source code, IP, meeting transcripts) before it can be sent to an external model.
Would have: Each of the three submissions would have been blocked or anonymized at the boundary — the leak becomes technically impossible, not merely against policy.
Provide an internal, data-controlled model for exactly these tasks (debugging, summarization) so the useful work happens inside the boundary.
Would have: Engineers would have gotten the help they needed without any data leaving Samsung's control.
Monitor and log AI-tool usage as a first-class egress channel in DLP, not an unwatched exception.
Would have: The first leak would have been detected immediately, before the second and third repeated it.
Why it wasn't caught
The AI-tool channel was outside existing data-loss-prevention monitoring, so nothing flagged confidential data leaving via ChatGPT. The leaks were discovered after the fact, by which point the data was irretrievable.
Recurrence risk
HIGH in any organization that permits external AI tools for real work without a technical egress control — policy and training alone don't prevent well-intentioned employees from pasting sensitive data.
Sources
Check your own agent against this failure pattern before it ships: