Pause and contain further exposure
If you notice that you have pasted sensitive information into an AI service, stop using that conversation for the task. Do not paste more private material to ask the tool how serious the mistake is. Avoid copying its response into another service, especially if the response repeats the information. A single mistaken submission is easier to assess than a chain of uploads and shares.
Check whether the conversation, file, or output has also been shared with other people. Look for share links, connected workspaces, exported chats, automated workflows, and documents where you may have copied the response. If you can disable a share link or restrict access, do so. Keep a brief record of what you changed and when; that record can help a team respond without relying on memory.
Work out what was shared and where it went
Identify the information as precisely as you can without creating another unnecessary copy. Was it a password, API key, customer name, private message, unpublished work, financial record, or internal document? Note whether it included one person’s details or a larger set, and whether the text alone could identify anyone. The potential response depends more on the content and who can access it than on the fact that an AI tool was involved.
Record the service and account you used, the approximate time, and how the information entered the tool: a typed prompt, uploaded file, browser extension, integration, or automated workflow. Check the service’s settings and documentation for the controls that apply to that account. History, retention, training-use settings, and deletion options can differ by product and account type. Do not assume that a private-looking chat is stored only on your device, or that deleting it removes every copy immediately.
Take action that matches the information
If you shared a password, API key, access token, or similar credential, treat it as exposed. Change or revoke it through the system that issued it, then check where the old credential was used so connected workflows do not fail unexpectedly. If the credential could access a team account, tell the person responsible for that system. Removing the prompt from the AI service does not make a disclosed credential safe to keep using.
For personal or confidential information, use any available controls to remove the chat, file, or share link, and contact the service’s support or privacy channel if you need clarification about removal or retention. Describe the submission without pasting the sensitive material again unless a secure process specifically requires it. These steps may reduce further access, but they cannot establish that nobody has already viewed, processed, or copied the information.
Notify the right people without widening the exposure
If the information belongs to an employer, client, collaborator, or customer, follow the relevant internal reporting process promptly. A manager, security contact, privacy lead, or project owner may need to coordinate the response. Give them the facts you have: what type of information was involved, which service received it, when it happened, who may have had access, and what you have already done. If you are unsure whether reporting is required, asking through the approved channel is safer than quietly guessing.
Be careful with screenshots and forwarded transcripts. They can turn a limited mistake into several new copies of the same sensitive information. Where possible, share a description first and use an approved secure channel if someone needs the exact content. Avoid promising affected people that the information has been fully erased unless you can verify that claim through the appropriate process.
Make the next use of AI safer
Once the immediate response is underway, identify how the mistake happened. Perhaps a document was uploaded without review, a prompt template invited people to include real records, or an integration sent more context than expected. Replace real names and identifiers with placeholders when the task allows it. For repeatable work, prepare a small set of fictional examples or an approved, redacted version of the material before opening the AI tool.
Create a simple pre-submission check for your own workflow or team: What data am I sending? Do I have permission to send it? Does this tool need the original details to complete the task? Who can access the conversation and its output? AI services can be useful without receiving every detail available. The goal is not to stop all use after a mistake, but to limit the next task to the information it actually needs.
