Samsung leaked confidential data into ChatGPT three times in 20 days

Three engineers in Samsung's semiconductor division pasted source code, an internal-meeting transcript, and chip yield-test sequences into ChatGPT. Twenty days later: a companywide ban. The fastest case study in why shadow-AI visibility comes before shadow-AI policy.

What happened

In April 2023, three separate Samsung Device Solutions (semiconductor) engineers used ChatGPT for productivity tasks within a 20-day window:

1. An engineer pasted proprietary semiconductor source code in to debug an issue.
2. A second pasted the transcript of an internal meeting in for summarization.
3. A third pasted test-sequence code used in chip yield optimization.

Samsung's policy didn't permit this and there was no monitoring layer. The company first imposed a 1024-byte prompt cap, then on May 1, 2023 banned all generative AI on company devices and internal networks, threatening dismissal for further violations.

The bigger pattern, in numbers

• Cisco 2025: 46% of organizations report internal data leaks via generative AI.
• UpGuard 2024: ~80% of workers (and ~90% of security pros) admit using unapproved AI tools at work.
• 2025 SaaS spend survey: 78% of IT leaders report unexpected SaaS or AI consumption charges, up from 65% in 2024.
• IBM 2025: only ~37% of organizations have any policy to detect shadow AI.

Samsung is the visible case. The unseen majority is where the long-tail damage compounds — IP exposure, GDPR-relevant flows, vendor-side breaches that retroactively turn private data into discoverable artifacts.

Why "ban it" doesn't actually work

Samsung's ban was rational but blunt. It assumed the company knew which AI tools were in use. In most enterprises, that assumption is false. Engineers with company credit cards can stand up an OpenAI account, an Anthropic account, a Cursor account, a v0 account in five minutes each. The IT org learns about it from the expense report a quarter later.

A ban without visibility is a policy without enforcement. Three more leaks happen on personal accounts run on company devices, and you find out from a Bloomberg article.

What would have changed it: visibility before policy

1. Inventory every AI provider account paid for by the company. An Admin API key from OpenAI, Anthropic, Azure OpenAI, Vertex/Gemini, or GitHub Copilot is the chokepoint — every account has one and it bills against your card.
2. Pull usage per team and per model. An Anthropic key billing $11k against the chip-design team is the signal you needed before incident #2.
3. Tag the keys. Each key has an owner, a team, a renewal date, a status (active / paused / revoked). When an engineer leaves, their keys revoke automatically.
4. Provide sanctioned alternatives. Once you know who is using what, you can stand up a governed proxy with logging, give teams a sanctioned prompt library, and convert "I'll just paste it into ChatGPT" into "I'll fetch the approved sentinel prompt."

This is the shape of Inventoria's AI Spend module. Single screen: every provider connection, every key, every team, every dollar. The control set isn't punitive — it's just visibility, the same kind your CMDB has had for hardware for fifteen years.

What you should do this quarter

1. Get every AI provider Admin API key into one place. Decommission unowned keys.
2. Map keys to teams. Bind teams to budgets. Bind budgets to alerts.
3. Publish your "approved alternative" — a sanctioned proxy or a prompt library — before you ban anything.
4. Reserve the ban for the cases where Step 3 doesn't move the needle. It's a much smaller cohort than you think.