What should we do if someone puts sensitive data in the wrong AI tool?
If someone pastes a password or client data into the wrong AI tool, here is the same day plan for a San Diego firm, including credential rotation, account checks and the IT setup around the tool.
Shane Donoher · Published November 19, 2026
The short answer
If someone has already put sensitive data into an AI tool the firm did not approve, treat it as an incident the same day. Tell the named person in the firm at once. Write down what was shared, in which tool, on which account and when. Delete the conversation where the tool allows it. If a password, API key or other credential went into the prompt, rotate it. Then ask your IT provider to check sign ins, mailbox rules and app permissions on that person's work accounts.
A recent vendor incident shows why the credential step matters. After a cyberattack at Fireworks AI, Notion published an advisory on 12 November 2026 for affected workspaces. Notion recommended that those customers treat the data as if it were exposed, and said: if people had pasted sensitive information, passwords, API keys or other credentials into Notion AI during the incident window, rotate them. (Notion security advisory, originally published 12 November 2026; TechCrunch, 12 November 2026) That advice is useful even when there is no public breach. Once text is in a tool you do not control, you should assume it may be retained, and act accordingly.
Why this comes up
Staff usually mean well. A personal account is already signed in on the phone. The approved tool is slow, or the person is working from home on a connection that cannot reach it. The prompt looks harmless until a client name, a patient detail, a password or a key is included.
The AI mistake and the IT setup are often the same problem. If the firm has no working business account, or the office network and firewall path make the approved tool unreliable, people will fall back to whatever is already open on their phone. We wrote earlier about what never goes in the prompt and why the plan on the account matters. This post is the next step: what to do after it has already happened.
What to do the same day
1. Tell someone immediately
Name one person who should hear about this, and a deputy if that person is out. Early reporting makes the rest of the work much easier. A staff member who comes forward the same day is helping, not causing the problem.
2. Write down the facts
Record, in writing:
- What was shared, in categories rather than by pasting the secret again. For example: a work password, an API key, a client name, a file, or a patient identifier.
- Which tool and, if you know it, which feature or plan.
- Whether it was a personal account or a work account, and whose.
- Which device was used, and whether it was on the office network, home Wi Fi, or a phone connection.
- The date and the approximate time.
Your IT provider cannot check the right sign in logs, devices or tenant settings without this. Your compliance adviser will need the same record if client, patient or tax data was involved.
3. Delete the conversation where the tool allows it
Sign in to the same account, delete the chat or document, and empty any trash if the product has one. If the tool was on a personal account, still do this, and then sign that account out of any work device. Deleting the visible conversation is a sensible step. It is not a guarantee that every copy the vendor has retained is gone. Ask the vendor, in writing, what is retained and for how long, if the data is sensitive.
4. Rotate anything that can be used to get in
If a password, API key, access token or similar credential went into the prompt, change it now and revoke the old one. Notion's advice after the Fireworks AI incident is a clear public example of this step. (Notion security advisory, 12 November 2026)
If the credential belongs to a work account, also:
- Confirm multi factor authentication is on, and that no new sign in method was added.
- Ask your IT provider to review recent sign ins, mailbox rules, forwarding and app permissions for that person.
- Check whether the same password was reused on any other work system, and change those too.
If the pasted material was client, patient or tax information rather than a login secret, rotation is not the main job. Keep the record, follow the firm's incident process, and talk to your adviser about whether anyone needs to be notified. That is their call, not a conclusion this post can make.
5. Ask IT to check the surrounding systems
This is where security and infrastructure earn their place in an AI incident.
- Review work email and the firm's cloud tenant for unusual sign ins, new devices, new mailbox rules and new app permissions.
- Confirm the unapproved tool is not still signed in on a work laptop or phone.
- Ask whether staff can actually reach the approved tool from the office, and from home, on a firm account. If the approved path is blocked by a firewall rule, a dropped connection or a missing licence, the next person will use the personal app again.
- Decide whether the unapproved site should be blocked on the office network and on managed devices, once a working approved alternative is in place. Blocking alone, with no approved option, usually pushes the work back onto a personal phone.
A walkthrough of the wider security basics is in cybersecurity solutions for small business in San Diego. For unsanctioned tools already in use, see how to find where firm data is going.
6. Close the gap
Once the immediate risk is handled, fix the reason it happened. Write a one line rule for what may go into an AI tool, name the approved tools, and put staff on business accounts the firm controls. Make sure those accounts, the devices and the network path actually work, so the approved route is the easy one.
What is still unknown
As of 19 November 2026:
- Notion has said that if you have not heard from them, they have not identified your workspace as affected, and that no action is needed on that incident right now. Their investigation was described as ongoing. Other firms should not assume they were, or were not, in any vendor incident unless the vendor has said so.
- We do not know, for a typical consumer or business AI tool, whether deleting a chat removes every retained copy. That depends on the vendor and the plan. Do not treat deletion as proof of erasure.
- We will not invent how often this happens in small firms. The practical response does not depend on a statistic.
We will update this post if Notion, Fireworks AI or another primary source publishes a material change that affects the steps above.
How we help
Managed Safe AI keeps approved tools, settings and the client data rule in place month by month, so a one off incident does not become the normal way of working. A Safe AI Setup moves staff onto business accounts with the right settings and writes the rule into your AI use policy. If you are not sure which tools are already in use, an AI Safety Check finds out first.
If the paste included a work password or you need sign in logs and mailbox rules checked, see our cybersecurity services in San Diego or Microsoft 365 services. Book a 15 minute call to talk it through.
FAQs
Is deleting the chat enough?
No. Delete it where the tool allows, because that is still worth doing, then rotate any credential that was pasted and check the person's work accounts. Deletion may not remove every copy the vendor has kept.
What if it was a personal account, not a work one?
Follow the same steps: record the facts, delete the conversation, rotate any credential, and sign the personal account out of work devices. You will have much less visibility than you would on a firm controlled account, which is one reason personal plans are a poor place for client or patient work. Move that person onto a business account the firm administers.
Do we have to tell the client, a patient or a regulator?
That depends on what was shared, who it concerned, and the rules that apply to your firm. This post cannot answer that for you. Talk to your own legal or compliance adviser the same day, with the written record from step 2 in hand.
Does this only apply after a vendor breach?
No. The same steps apply to an ordinary mistake, such as a prompt pasted into a personal chatbot. A public vendor incident is only one reason the text may later be exposed.
What if the prompt only had a first name or a public fact?
Still tell the named person and write it down. The follow up will usually be lighter than a pasted password or a client file, but the approved tool rule should still be restated so it does not drift.
---
*This post is general information, not legal, tax or compliance advice. We checked Notion's security advisory and TechCrunch's 12 November 2026 report on 19 November 2026. Vendor terms, retention practices and professional duties can change, so check the current position and talk to your own adviser about your firm's situation.*