When personal data leaks

Why this matters

Sooner or later, personal data will leave through a place it should not. A well-meaning officer pastes a case file into a public chatbot. A file is sent to the wrong person. A tool is set up with the wrong sharing settings. What matters then is how fast the office learns of it and how calmly it responds. Most of the damage from a leak can be reduced if it is found early.

How leaks happen with AI

• Pasting or uploading personal data into a tool the department has not approved. This is data leakage, and when the tool is unapproved it is shadow AI.
• Wrong settings, such as a shared link that anyone can open or a feature that keeps conversations.
• Tricked tools. A tool that reads documents or web pages can be steered by hidden instructions in them. This is prompt injection. It is one reason why a tool with access to many files needs careful limits.
• Provider problems. The provider itself suffers a breach or changes its terms.
• Ordinary mistakes. The wrong attachment, the wrong recipient or a screen shared in a meeting.

The first hour

Follow your department's procedure and the law, including any duty and deadline to report. This course does not set any. In the absence of a better plan, a sensible order is:

1. Stop. The person who made the mistake stops using the tool and does not try to quietly fix it alone.
2. Tell. Report at once to the named officer, such as the privacy officer or the head of IT. Speed matters more than a complete report.
3. Contain. Remove access where you can: delete the shared link, close the account, withdraw the file and ask the provider to delete what was sent. Keep in mind that deletion at a provider may not be certain.
4. Record. Write down what happened, when, which data, how many people, which tool and what has been done so far. Keep screenshots and messages. This is your audit trail.
5. Assess. Ask who could be harmed and how. A leaked phone number is different from a leaked medical record. Decide which people, if any, must be told and by whom.
6. Fix and learn. Find the cause, close the gap and tell the office what changed.

A culture where people report

People report quickly only when they expect to be treated fairly. If every mistake brings punishment, mistakes get hidden and small leaks become large ones. Make it plain that reporting at once is the right thing to do, and treat concealment as the serious matter. Often a leak through an unapproved tool means the office has not given staff a good approved option. Our guide to shadow AI in your team looks at this.

Prepare before it happens

• Publish one short page: who to tell, how and what to include.
• Keep an up-to-date list of approved tools and the settings they must use.
• Make sure the first person told knows what to do next.
• Practise once a year with an invented incident.

Try it

Invent a leak: an officer pastes 20 complaint letters, with names and phone numbers, into a free chatbot. Write the one-page incident note covering what happened, what data, who is affected, what you did in the first hour and what you would change.

Check yourself

1. Why should the person who made the mistake tell someone straight away, rather than fix it quietly?
2. Name three things to write in the incident record.
3. Why might a leak through an unapproved tool also point to a gap in the office?

Answers: (1) Early reporting gives the office the best chance to limit harm and quiet fixes hide the extent of it. (2) Any three of: what happened, when, which data, how many people, which tool, actions taken. (3) Staff may have turned to it because the office has not given them a good approved option.