AI and Company Confidentiality: What Not to Give a Chatbot

To write a report, an employee uploads a sales spreadsheet full of customer data into a chatbot. To find a bug in a script, a colleague pastes an entire server configuration file, passwords included. Neither means any harm, but in both cases confidential company data has ended up on an outside service’s servers, where you have no control over it. This is the core privacy problem with AI: not the technology itself, but the data that flows into it without a second thought.

What happens when you give text to a chatbot?

With cloud services, your text and files are sent to the provider’s servers and processed there. Depending on the account type and terms of use, the following may happen:

• Conversation history is kept for a set period or until the user deletes it.
• Conversations are used to improve the model, which is the default in some services and can be turned off.
• Human review of some conversations for quality control or abuse detection.
• Connections to other tools such as email or cloud storage, which widen the chatbot’s access.

These terms vary a lot between free, personal and business accounts, and providers change them over time. Don’t rely on what was true last year; read the current terms.

What information should never be entered?

• Passwords, API keys, tokens and certificates, even inside a configuration file or log.
• Identity and contact details of customers and employees, bank account numbers and personal financial data.
• Contracts, pricing and negotiation documents whose disclosure would harm the company.
• Network diagrams, internal addresses and firewall settings, which are a roadmap for an attacker.
• Proprietary source code of company software and details of unreleased products.

Before entering any text, ask yourself: if this text ended up with a competitor or the public tomorrow, would it cause a problem? If the answer is yes, don’t enter it.

How to keep using AI safely

1. Sanitize the data: before sending, replace names, numbers, IP addresses and passwords with placeholder values such as CUSTOMER_1 or 10.0.0.x.
2. Send only what is needed: to find a bug in a script, the few problem lines are enough, not the whole file.
3. Use a business account: business editions usually give more control over data retention and use; check the details in the service agreement.
4. Review privacy settings: check the option for using your data to train the model and how long history is kept, and set them according to company policy.
5. For highly sensitive data, consider a local model: some open-source models can run on a server inside the company so data never leaves the network. On the other hand, this requires suitable hardware, maintenance and server security, and output quality may differ from cloud services.

The role of the IT team

Policy without technical controls is not enough. The IT team can define a list of approved services, require sign-in with business accounts and two-factor authentication, restrict access to unapproved services at the firewall or proxy where possible, and use data loss prevention (DLP) tools to monitor sensitive data being sent out. Short, regular staff training is also an important part of the job; most leaks come from lack of awareness, not malice.

Checklist

• The list of data that must not be entered into AI services is written down and communicated.
• Approved services and account types are defined.
• Data retention and usage terms of the services have been reviewed and are checked periodically.
• Employees know how to sanitize data before sending it.
• Accounts are protected with two-factor authentication.
• For highly sensitive data, the option of running models locally has been evaluated.

Protecting data in the age of AI is a continuation of the same old information security principles. ITPlus724 engineers can support you with firewall configuration, access control, network security and preparing in-house servers to run these tools locally.