What to Do After an AI Data-Sharing Mistake

Report the facts, even while the impact is unknown

You realise you uploaded a client attachment to your personal AI account. Or an assistant connected to a wider folder than intended. You do not yet know who could access the information or whether it was retained.

Treat that uncertainty as a reason to report promptly through your organisation's incident process. Do not wait for proof that someone misused the data. Your job is to explain what happened and prevent further exposure within your authority; the response team assesses impact and directs the next steps.

The first actions

1. Stop the activity. Do not send more prompts, repeat the upload or forward the generated output. Stop a running action using the approved controls if you can do so safely.
2. Contact the designated channel. Use the security or IT incident route and involve the data owner as the process requires. If you do not know the route, use a known IT contact or manager to locate it promptly. Use an alternative known channel if the usual one may be compromised.
3. Record a short timeline. Note the time and timezone, service and account type, feature, affected file reference, data categories and actions taken. Separate confirmed facts from things you do not know.
4. Preserve relevant evidence through the approved process. Do not erase records or distribute new copies while trying to fix the mistake. Let the response team direct deletion, retention, provider contact and other containment steps.
5. Follow the assigned response. If a password or access key may be exposed, alert the responsible team urgently so it can revoke or rotate it through the authorised process. Do not try to hide the exposure by only changing the prompt.

The NCSC incident response guidance emphasises defined roles, escalation, containment and records of response actions. Your organisation's plan determines who contacts the provider, affected people, customers or regulators. Do not make promises about the outcome or invent a legal reporting deadline.

A useful incident note

What happened: [factual description]
When: [time, date and timezone; say if approximate]
Tool/account/feature: [service, work or personal account, upload or connection]
Affected information: [categories and approved file reference; no passwords or full records]
Scope: [known files, recipients or actions; unknown items clearly marked]
Actions already taken: [what, when and by whom]
Evidence location: [approved location, not a public sharing link]
Help needed: [containment and next steps]

I have stopped the activity and have not confirmed retention or external access.

Do not paste the exposed data into another chatbot to ask whether it was serious. Do not attach the full file to a broadly visible ticket if an approved reference is enough. Let the response team choose how to obtain the material it needs.

Why deletion is not an all-clear

Deleting a visible chat may be one step in an authorised response, but you cannot infer that it removed every log, retained copy or connected-service record. Likewise, turning off future model training does not establish what happened to an earlier transfer. Record what you actually did and what remains unknown.