In the spring of 2023, a handful of engineers at one of the world's largest technology companies did something entirely ordinary. They used a new AI chatbot to help with their work. Within weeks, the company had banned the tool. The Samsung incident became the reference case for ChatGPT data leakage, and it still contains the clearest lessons for any company deciding what may and may not go into a public AI tool.
What happened at Samsung? A reconstruction from news reports
The facts below come from published reporting. Samsung did not publicly confirm the detailed accounts, so treat them as reported, not established.
The three reported incidents
According to Engadget, citing The Economist Korea, three employees in Samsung's semiconductor business leaked confidential information to ChatGPT soon after the division began allowing engineers to use it:
- An employee submitted sensitive database source code to check it for errors.
- Another submitted code and asked for it to be optimised.
- A third submitted a recorded meeting to generate minutes.
Engadget reported that Samsung responded by capping the length of ChatGPT prompts, investigating the employees and beginning work on its own internal chatbot. The Register later reported, from a memo seen by Bloomberg, that Samsung told staff in one of its biggest divisions not to use generative AI tools such as ChatGPT, warning that violations could lead to termination, and that the reported leaked data included equipment measurement and yield data from its semiconductor business.
Why it counts as a leak even without a hacker
No attacker was involved. Nobody broke in. The data left because employees sent it to a third-party service. Once confidential text sits on someone else's servers, you are relying on that provider's retention, access controls, staff and terms of service. The Register noted that at the time, OpenAI's policy was that data submitted through its API was not used for training unless the customer opted in and was kept for up to 30 days for abuse monitoring, and that a setting to stop ChatGPT conversations being used for training had only just been introduced. Consumer and API services can follow different rules, and the rules change. That uncertainty is the point.
The root cause was not careless engineers
It is easy to blame the individuals. That misses three structural problems most companies still have:
- Access arrived before rules. Staff were allowed to use the tool before anyone had told them what data was off limits.
- The tool was genuinely useful. Debugging code and writing minutes are exactly the jobs AI does well. People will reach for it.
- Nobody had classified the data in terms an employee could apply in five seconds. "Don't share confidential information" is too vague when the person holding a stack trace does not think of it as confidential.
The rule: what may and may not go into a public AI tool
Turn the lesson into a rule that fits on a desk card. We suggest three colours, because people remember colours under deadline pressure.
| Colour | What it covers | Public AI tools (free or personal accounts) |
|---|---|---|
| Green | Anything already public or that would cause no harm if published: public website text, general how-to questions, open-source code, anonymised examples you wrote from scratch | Allowed |
| Amber | Internal but low-impact material: internal process descriptions, generic code without business logic or identifiers, drafts with names and numbers removed | Only through a company-approved AI tool on a company account |
| Red | Source code with business logic, configs, logs, database schemas, credentials, customer or employee data, contracts, pricing, financials, board or strategy material, meeting recordings, anything under NDA | Never. Use only a tool approved for confidential data, such as a private deployment |
Map each Samsung incident to the rule
- Proprietary source code for debugging: Red. Write a minimal example that reproduces the bug without proprietary logic, or use an approved private tool.
- Code for optimisation: Red if it encodes how your process or product works, which in a semiconductor fab it almost certainly does.
- Meeting recording for minutes: Red. Meetings contain names, decisions, numbers and plans. Use an approved transcription and summarisation tool with known data handling.
Five questions to ask before you paste
Give staff a quick self-check. If the answer to any of these is yes, the content is Red:
- Would I be uncomfortable if this text appeared on a public website tomorrow?
- Does it contain any person's name, phone number, email, ID number, address, health or financial details?
- Does it contain a password, key, token, internal hostname or IP address?
- Does it describe how our product, process or pricing works in a way a competitor would value?
- Is it covered by a client contract or NDA?
Common leakage paths beyond copy-paste
The Samsung reports involved direct pasting. In 2026 there are more routes, and your rule should cover them:
- File uploads: PDFs, spreadsheets and slide decks uploaded for summarising.
- Browser extensions that read every page you open, including internal dashboards and web mail.
- Meeting bots that join calls, record them and send transcripts to an external service.
- Coding assistants that send surrounding code, and sometimes configuration files, as context.
- Connected apps that are given access to mailboxes, drives or chat workspaces through a single "Allow" click.
Technical controls that back up the rule
- Single sign-on for approved AI tools, so use is tied to company accounts with admin logs and correct data settings.
- Data loss prevention rules that detect sensitive patterns, such as keys, card numbers, PAN or Aadhaar formats, before they are pasted or uploaded to web services.
- Secret scanning in repositories and CI, so credentials are not lying around to be pasted in the first place.
- Extension allow-lists on managed browsers.
- A private AI option for Red data, so engineers and analysts have a safe place to do the work they would otherwise do in a public tool.
What Samsung's response teaches
Look at the reported sequence: allow the tool, suffer leaks, cap prompts, ban the tool, build an internal one. The last step is the real fix, and it could have come first. If you want staff to stop pasting sensitive data into public AI, give them an approved tool that is just as easy to use for Red data, and make the rules simple enough to remember.
For Indian companies, the stakes include the Digital Personal Data Protection Act: Section 8(5) requires a data fiduciary to keep reasonable security safeguards against personal data breaches, as Bar & Bench explains, so customer and employee data pasted into an unapproved tool remains your problem. A separate article on shadow AI on the Cehpoint blog covers how to discover which AI tools are already in use and how to roll out a sanctioned-AI policy.
How Cehpoint can help
Our private LLM deployment service sets up an AI assistant inside your own cloud account or data centre, so source code, contracts and customer data can be worked on with AI without leaving infrastructure you control. Our IT consulting team can write your data classification rule and AI-use policy, and align them with ISO 27001 and DPDP requirements.