How to Respond to an AI Data Leak in the First 72 Hours
Someone pasted customer data into a chatbot, or an agent printed an API key into a log. The first hours decide how bad it gets, and the order of steps matters.
To respond to an AI data leak, do these in order: contain it so it stops getting worse, preserve the evidence, work out what was exposed and to whom, then decide who has to be told and by when. Under the GDPR, a controller has to notify the supervisory authority without undue delay and, where feasible, within 72 hours of becoming aware of a breach that is likely to put people at risk.
This is a practical sequence, not legal advice. The rules differ by country and by the kind of data involved, so bring in a lawyer or your data protection contact early. What follows is what to have moving in the first hours so that conversation starts from facts.
First, identify which kind of leak this is
AI tools tend to leak data in three ways, and the first move differs for each.
Type | Example | First move |
|---|---|---|
Pasted in | An employee pastes customer records or a contract into a chatbot | Find which tool and account, check its data retention and training settings, and request deletion where the provider allows it |
Output exposed | An agent or app prints a secret or personal data into a log, a commit, a chat reply or a public page | Take the exposed copy down and rotate any credential in it immediately |
Vendor side | The AI vendor reports an incident, or your data turns up where it should not | Get the vendor's incident details in writing and establish exactly which of your data was involved |
Do not spend an hour debating the category. A rough label is enough to pick the first move, and you can correct it as facts arrive.
Hour 0 to 4: contain
The FTC's data breach response guide opens with securing operations: act fast to contain the breach and fix its cause, because "the only thing worse than a data breach is multiple data breaches".
Rotate credentials first. The guide notes that a stolen credential keeps working until it is changed, even after the attacker's tools are gone. If an API key, token or password appeared anywhere it should not have, revoke and reissue it now. This is the single highest-value action in an AI leak, since agents and prompts are a common place for keys to end up.
Remove exposed copies. The FTC guide advises taking down personal data posted in error and asking search engines and other sites to remove cached or copied versions.
Stop the source. Pause the agent, disable the integration or revoke the tool's access to the data store while you investigate.
Do not delete evidence. Take down what is public, but keep logs, chat histories, screenshots and timestamps. The guide tells you to document the investigation and preserve forensic evidence.
Hour 4 to 24: scope it
Answer four questions in writing. A shared document with timestamps is enough.
What data? Names, emails, payment details, health information, credentials, source code, contracts?
Whose data? Customers, employees, a client's customers? Roughly how many people?
Where did it go? Which tool, which vendor, which account, which public page?
For how long? When did it start and when did it stop?
The FTC guide asks you to work out the same things: what was compromised, how many people are affected and whether you have their contact details. If the leak went into a vendor's system, read its terms to see whether the data is retained or used for training. Our guide on checking whether an AI tool trains on your data shows where to look, and vetting an AI vendor covers what to ask them.
If you run an AI-built app, an audit log is what makes these questions answerable. Without one, you are reconstructing events from memory.
Hour 24 to 72: decide who must be told
Regulators under the GDPR. Under GDPR Article 33, the controller must notify the supervisory authority "without undue delay and, where feasible, not later than 72 hours after having become aware of it". A late notification must include reasons for the delay. No notification is needed if the breach "is unlikely to result in a risk to the rights and freedoms of natural persons". Information can be given in stages when it is not all available at once.
The notification itself must at minimum describe the nature of the breach including approximate numbers of people and records, give a contact point such as a data protection officer, explain the likely consequences, and describe the measures taken or planned.
If you are a processor, meaning you handle data on behalf of a client, Article 33 requires you to alert the controller without undue delay once you become aware, and the 72-hour window is theirs, not yours. Tell your client right away.
In the United States, the FTC guide notes that all states, DC, Puerto Rico and the Virgin Islands have breach notification laws, and that other federal rules may apply, including for health data. Your legal obligations depend on where the affected people live, so find out before you decide there is nothing to report.
Affected individuals. The FTC guide advises notifying them quickly and saying what happened, what information was involved, what you are doing about it and what they can do. Designate one point of contact. Where Social Security numbers or financial data were involved, consider offering credit monitoring and tell people how you will contact them, so they can spot phishing that exploits the incident.
Your customers and partners. If you hold their data as a processor, or the leak touched something they gave you, tell them before they read about it elsewhere.
Then fix the cause of the AI data leak
Containment is not the end. Close the path that led to the leak:
For pasted-in data, set a clear rule on what may go into which tools, and provide an approved tool so people are not tempted to use a personal account. Our guide to sandboxing an AI agent covers limits for the agent side.
For output exposure, keep secrets out of prompts and environments the model can read, and scan logs and commits for credentials.
For vendor incidents, review the contract and decide whether you can keep using the product.
For prompt-injection routes, where an attacker's text makes your tool disclose data, see what prompt injection is and apply input and output controls in your app.
Write a short review within a couple of weeks: timeline, cause, what changed. The broader landscape is in the AI risks guide.
FAQ
How long do I have to report a data breach under GDPR?
The controller must notify the supervisory authority without undue delay and, where feasible, within 72 hours of becoming aware of it, unless the breach is unlikely to result in a risk to people. Late notices must explain the delay.
Is pasting customer data into ChatGPT a data breach?
It can be, depending on the data, your agreements with the provider and the rules that apply to you. Treat it as a possible incident, find out where the data went and ask your legal or data protection contact to assess it.
What should I do first if an API key was exposed?
Revoke it and issue a new one immediately. A stolen credential keeps working until it is changed. Then check logs for use of the old key.
Do I have to tell customers about an AI data leak?
It depends on the data and the law where they live. The FTC guidance advises notifying affected people quickly where their information was exposed, with clear steps they can take. Confirm your specific obligations with counsel.
How did this land?
About the author
Senior Editor, AI & Product
Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.
