ChatGPT Data Leak Prevention: What Samsung's 20-Day Lesson Means for Every Enterprise

Samsung Semiconductor learned an expensive lesson in 20 days. On March 11, 2023 — three weeks after lifting its internal ChatGPT ban — an engineer pasted proprietary semiconductor measurement database source code into ChatGPT, asking the model to identify bugs and optimise the code. Within the same month, two more incidents followed: a second employee fed equipment defect detection algorithms into the chatbot, and a third recorded a confidential internal meeting, pasted the full transcript into ChatGPT, and asked it to generate meeting notes.

By May, ChatGPT was banned on all Samsung company devices and networks — permanently. An internal memo warned employees that OpenAI servers retain submitted data, and that once proprietary information enters a public LLM, it cannot be retrieved or deleted. The total cost of the breach was zero dollars in fines and infinite in exposure. (Sources: Forbes, CIO Dive, The Economist Korea, May 2023)

Samsung's case made headlines. But four years later, the behaviour it exposed has not slowed down. It has accelerated.

The Numbers That Should Keep Your CISO Awake

Cyberhaven's 2023 study, analysing 1.6 million knowledge workers, delivered a statistic that should appear in every board deck that touches AI procurement: 11% of data employees paste into ChatGPT is confidential. That includes proprietary source code, client records, regulated financial data, and internal strategy documents. Only 4.7% of employees had pasted confidential data at all — but the concentration was extreme: less than 1% of workers were responsible for 80% of all incidents.

Cyberhaven's 2024 follow-up study, now covering 7 million workers, showed the problem had not plateaued — it had nearly tripled. 27.4% of corporate data pasted into AI tools now qualified as sensitive, up from 10.7% a year earlier. Of the AI tools monitored, 71% were classified as high risk.

LayerX Security went further. Their analysis found that 77% of employees have shared sensitive company data via AI tools — not just ChatGPT, but Claude, Gemini, Copilot, and dozens of lesser-known AI SaaS products that procurement teams have never reviewed.

Harmonic Security's 2025 study identified sensitive information in over 4% of prompts to AI tools — meaning roughly one in every 25 prompts contains data that would trigger a disclosure obligation under regulations like the EU AI Act or GDPR.

What Actually Gets Leaked: The 8 Data Types

Consolidating findings from Cyberhaven, LayerX, Harmonic Security, and independent incident databases, eight categories of data are flowing into public LLM interfaces daily:

1. Proprietary source code — engineers debugging in ChatGPT, Claude, or Copilot with live production code
2. Client PII and database schemas — customer names, addresses, and table structures pasted for "format this for me" tasks
3. API keys and internal trade secrets — authentication tokens and proprietary algorithms
4. Internal financial data and projections — FP&A teams using AI to format board reports
5. Confidential meeting transcripts — the Samsung pattern, repeated weekly across thousands of organisations
6. Patient medical records — healthcare workers using AI for note summarisation
7. Patent-pending R&D materials — research teams accelerating innovation at the cost of IP protection
8. Equipment and algorithm specifications — manufacturing and engineering specifications exposed in search of optimisation

The pattern is consistent: employees are not malicious. They're trying to work faster. The tool that accelerates their workflow happens to be a public data ingestion pipeline they don't understand.

Why Traditional DLP Doesn't Catch This

Data Loss Prevention tools were architected for email attachments, USB drives, and file transfers — channels with predictable protocols and inspectable payloads. Copy-paste into a browser-based LLM interface bypasses these entirely. The data moves from clipboard to DOM to OpenAI's API in milliseconds, with no file attachment, no email gateway, and no endpoint agent triggering an alert.

Secuvy.ai's analysis of the "Copy-Paste Leak" pattern is blunt: "It is almost impossible to catch with traditional Data Loss Prevention." The data never touches a monitored channel. It goes straight from an employee's clipboard to a third-party server in a TLS-encrypted stream that looks identical to legitimate web traffic.

The Enterprise Response: Detection Before Policy

Policy documents are necessary but insufficient. Samsung had an AI usage policy before the three incidents. Every enterprise with an employee handbook has language about protecting confidential information. The gap is detection: you cannot enforce a policy against behaviour you cannot see.

The intervention points are threefold:

1. Spend-level detection — Unauthorised AI tool subscriptions appear on corporate cards and expense reports under innocuous merchant descriptors. ChatGPT Team seats show up as "OPENAI * CHATGPT SAN FRANCISCO CA." Midjourney subscriptions bill as "DRI* MIDJOURNEY.COM." Without transaction-level signature matching, finance teams miss them entirely.

2. Network-level monitoring — Browser extensions and endpoint agents that detect text pasted into known LLM domains can flag incidents in real time. Cyberhaven, LayerX, and similar providers operate in this space. The challenge is coverage: these tools only work on managed devices, and the average enterprise now runs 2,191 SaaS applications (Zylo 2026 Index), only 15% of which are controlled by IT.

3. Procurement-level governance — The most durable fix is making AI tool procurement visible before it becomes a detection problem. If the Office of the CFO can see every AI SaaS subscription across every department — not just the ones IT approved — the Shadow AI surface area shrinks to a manageable perimeter.

The AuditSentinel Position

AuditSentinel's discovery engine scans corporate transaction data for AI tool signatures — ChatGPT, Claude, Midjourney, Copilot, Gemini, and 60+ other AI SaaS products — surfacing unauthorised subscriptions that policy documents alone cannot catch. The platform's risk classification engine maps each detected tool against your organisation's compliance posture, identifying data sovereignty exposure before it becomes a breach notification.

The Samsung lesson was not that employees are careless. It was that governance systems designed for an era of on-premise software cannot protect data in an era of browser-based AI. Detection is not optional. It is the prerequisite for enforcement.

This briefing was prepared by the AuditSentinel Governance Intelligence Unit. Sources: Cyberhaven 2023 & 2024 studies, LayerX Security, Harmonic Security 2025, Forbes, CIO Dive, The Economist Korea.