Disaster 2: Samsung, early 2023 — engineers pasting source into ChatGPT

Spring 2023. Three weeks after Samsung Semiconductor lifted an internal ban on ChatGPT, three separate engineers leaked proprietary information by pasting it into the chat window:

1. An engineer pasted source code for a confidential semiconductor program into ChatGPT to ask for help debugging it.
2. A second engineer pasted internal test sequence code, asking ChatGPT to optimize it.
3. A third engineer pasted a recorded meeting transcript and asked ChatGPT to summarize it into minutes.

Each paste was a polite, well-intentioned use of a productivity tool. Each paste sent confidential corporate IP through OpenAI's API to be processed on their servers. At the time, OpenAI's default terms allowed conversation data to be used for model training. The engineers had read no policy that told them not to. ChatGPT felt like a private notebook. It wasn't.

What actually failed

The pre-2023 model of corporate security was built around a boundary: data is "inside" the company or "outside." Email attachments going outside got scanned. File-share access got audited. USB ports got locked. ChatGPT broke this model in a single keystroke: paste → submit → outside. The paste happened inside the browser, on the corporate laptop, on the corporate network. Every existing DLP control assumed the threat was deliberate exfiltration. Pasting into a help tool is the opposite — it's an engineer trying to do their job faster, with zero adversarial intent.

The chain that failed:

1. No policy explicitly named ChatGPT as out-of-bounds. The ban had been lifted three weeks prior. Engineers assumed lift = approved.
2. No DLP rule flagged "this paste contains code from an internal repo." The corporate DLP suite knew what was in email, in Box, in Salesforce. It had no hook into a browser paste event.
3. No enterprise plan with opt-out from training. The free ChatGPT consumer plan defaulted to using conversations for training. Enterprise plans (with data-use opt-outs) existed but Samsung hadn't bought one.
4. No internal AI tool engineers could use instead. Samsung had not yet built or licensed a private LLM. The engineers had no sanctioned way to do the same task.

The third failure is the load-bearing one. If you tell engineers "don't use AI for help" but give them no sanctioned alternative, they will use AI for help. The policy is theater.

How this connects to git

You're in a git lesson, so why this incident? Because the Samsung paste is the same failure mode as committing a secret to a repo, expressed through a different surface:

Surface	Disaster pattern
git add + git commit + git push	Secret in version control. Public if repo is public; recoverable if repo gets breached.
Paste into ChatGPT consumer plan	Source / data leaves your perimeter, may be retained, may train the next public model.
Paste into a Slack channel	Source / data archived in Slack indefinitely, accessible to anyone with channel access.

In every case, the developer trusted a tool to be a private notebook when it was actually a publishing surface. The git version of this is the one you'll hit first as a builder, but the pattern repeats everywhere AI tools sit in the workflow.

This becomes the dominant pattern in chapter 18 (secrets and env management), which is why this lesson sits right before it. The git workflow is one of many places the same mistake can land.

The control chain that would have caught it

In order from earliest to latest:

• An enterprise plan with training opt-out, configured by default in the corporate Chrome profile. If the paste happens but the data doesn't train the model, the long-tail exposure shrinks dramatically. Doesn't prevent the paste; it caps the blast radius.
• A browser-extension DLP rule that intercepts pastes containing patterns matching internal source (proprietary class names, internal package prefixes, the literal phrase "Samsung Confidential"). Catches the paste at the moment of attempt. Annoying for engineers; the cost of the annoyance is measured against the cost of the breach.
• An internal AI assistant trained on internal docs, available via an internal URL, with a clear "use this, not ChatGPT" policy. Removes the engineer's motivation to reach for the consumer tool in the first place.

The third is the most expensive but the most effective. Samsung accelerated their internal tool program after this incident.

What got rotated

Less "rotated" and more "policy-and-process changed":

• ChatGPT use was re-banned internally, this time with monitoring.
• An internal LLM project was put on a six-month delivery clock.
• The three engineers were disciplined (no public detail on how).
• A broader DLP review was opened across the semiconductor unit.

The leaked source code itself is, in principle, in OpenAI's training data. Or it isn't. Nobody outside OpenAI can prove either way. That's the part of this breach you can't undo and can't audit. The most damaging consequences are the ones you can't see.

What you should take from this

When an AI tool is in your workflow, the question is never "do I trust this tool?" — it's "where does this data go after I hit enter?" For ChatGPT consumer plan, the answer was "OpenAI's training set." For Cursor, the answer is whatever's in their data-use policy this quarter. For Claude Code, the answer is Anthropic's policy. For your company's internal tool, the answer is "the same place as our other internal data."

Don't paste anything into anything until you can answer that question. That's a habit. Build it now.