Insights

An AI vendor breach exposed Notion AI chats. What it means for the AI tools your team uses

Fireworks AI's 17 November update shows how one AI vendor breach reached Notion and Uber customers. Here is how a San Diego firm can check which AI vendors touch its data and lock down access.

Shane Donoher · Published November 20, 2026

The short answer

On 17 November 2026, Fireworks AI, a platform other software companies use to run AI models, published its latest update on a breach that reached customers of those companies. Notion AI chat transcripts from under 1% of Notion workspaces were taken, and Uber has told close to 150,000 drivers that their licence details were stolen. Most of the people affected had never signed up with Fireworks. They used a product that relied on it.

For a 10 to 100 person firm in San Diego, the useful lesson is simple. Every AI tool your team uses sits on a chain of other vendors, and your client or patient data travels as far as that chain goes. This week, list the AI tools that touch firm data, find out which outside vendors each one passes that data to, and make sure each tool signs in through your Microsoft 365 identity, can reach only what it needs, and can be cut off in minutes.

We covered what to do after sensitive data goes into the wrong AI tool yesterday. This post is the step before: choosing and checking the vendors so a breach two companies away does less harm.

What happened, with dates

  • 9 and 10 October 2026. Fireworks published its first bulletin. Uber later said in state notices that its provider told it on 9 October that hackers had stolen internal credentials and used them to reach Uber data. (Fireworks AI security bulletins; TechCrunch, first published 13 November 2026, updated 16 November)
  • 2 to 4 November. Fireworks told Notion that data had been accessed. Notion disconnected its systems from Fireworks on 3 November and deleted its data held there on 4 November. (Notion security advisory)
  • 8 and 9 November. Fireworks confirmed the affected dataset, which Notion found included Notion AI chat transcripts used for its Insights feature. Notion says most of those transcripts could not be tied to a person or workspace. Notion contacted affected customers on 9 November. Fireworks said the same day that the attackers appear to be a financially motivated group selling stolen access to large language models, a scheme known as LLMjacking. (Notion security advisory; Fireworks AI, 9 November 2026)
  • 11 to 13 November. Notion revoked Notion API keys it found in affected messages, reported other companies' API keys found in those messages to the vendors that issued them, and told affected customers to rotate any passwords, keys or credentials pasted into Notion AI during the incident window. (Notion security advisory)
  • 13 to 16 November. Uber's notices put the total at 149,346 people, with names, addresses, dates of birth and driver's licence numbers taken. Uber said it had alerts on its Fireworks environment, but because the attackers used Fireworks' own internal credentials, no alerts were triggered. (TechCrunch, updated 16 November 2026)
  • 17 November. Fireworks now describes two separate incidents. The first ran from 3 August to 7 October and used stolen Fireworks credentials and a flaw in one of its services. The second ran from 23 to 30 October, when attackers reached Fireworks development machines through a cloud provider that had its own security incident, and then reached part of production. Account and billing information for all accounts created before 30 October was accessed, plus datasets, models or secrets for a small number of customers. Fireworks says it does not store payment card numbers and that both incidents are contained. (Fireworks AI security bulletins, 17 November 2026)

The same pattern showed up again this week outside AI. On 19 November, JPMorgan Chase, Visa and Mastercard each said outside service providers had been hit, after an extortion group leaked customer names and email addresses. None of the three has named the vendors. (AP, 19 November 2026; SecurityWeek, 19 November 2026)

Does this affect your firm?

If your firm uses Notion and has not heard from Notion, Notion says it has not identified your workspace as affected and no action is needed now. If you were contacted, follow Notion's instructions and our same day steps.

The wider point applies to every firm using AI. A note taker, a writing assistant, a document tool or a chatbot built into software you already pay for will often send text to another company that runs the model, stores logs or provides the cloud. Vendors usually call those companies subprocessors. You may have accepted them in terms of service that nobody in the office read.

In our experience, the firms that come through a vendor incident calmly are the ones that already know three things: which tools hold their data, who signs in to them, and how to cut them off. Each of those is an identity, security or infrastructure question as much as an AI one.

Why the vendor chain is an identity and security problem

The Uber detail deserves a second look. Uber had monitoring on its vendor environment and it stayed quiet, because the attacker arrived with the vendor's own valid credentials. Alerts will rarely see a breach inside someone else's company, so limit what each vendor can reach in the first place, and make sure you can revoke it fast.

For a small firm on Microsoft 365, that comes down to a few settings and habits:

  • Sign in through the firm's identity. An AI tool that staff reach with their Microsoft 365 work account, with multifactor authentication and Conditional Access, can be switched off for everyone in one place. A tool signed up with a personal email or a separate password is invisible to the firm and keeps working after someone leaves.
  • Check which apps have been granted access to mail and files. Many AI tools ask for permission to read a mailbox, a calendar or SharePoint. In the Microsoft Entra admin centre, the enterprise applications list shows which apps users or admins have approved and what they can reach. Requiring admin approval for new app consent stops a staff member from handing a new AI vendor access to the whole mailbox with one click.
  • Keep secrets out of chats and documents. Notion found API keys inside chat transcripts. Passwords and keys belong in a password manager or a proper secrets store, never pasted into an AI prompt or a shared page.
  • Know what leaves the network. Your firewall or DNS filtering logs show which AI services the office actually talks to. That list is often longer than the list of approved tools, and it is the honest starting point for any review. Once an approved tool works reliably over the office internet and Wi-Fi, you can block the unapproved ones without people working around the block.

The Microsoft side of this, including the new controls for AI agents announced at Ignite, is in our post on which AI agents your staff installed. For a broader look at unsanctioned tools, see how to find where firm data is going.

What to do this week

  1. List the AI tools that touch firm data. Include AI features inside software you already use, such as document, note taking, practice management and accounting tools. For each one, write down who owns it in the firm and what data goes in.
  2. Find each vendor's subprocessor list. Most business software vendors publish one or will provide it on request. Look for the companies that run the AI model, store logs and host the service. Sign up for change notices where the vendor offers them.
  3. Check retention and training settings. Find out how long prompts and outputs are kept, whether they are used to train models, and whether any optional feature, such as analytics or insights, sends content to another company. Turn off what you do not need.
  4. Move sign in onto Microsoft 365. Where the plan allows single sign on, use it. Where it does not, make sure the account is a firm account with multifactor authentication, owned by the firm, and on the offboarding checklist.
  5. Tighten app consent in the tenant. Review enterprise applications in Entra, remove AI apps nobody owns, and require admin approval for new consent requests.
  6. Read the contract for breach notice. Check how quickly the vendor must tell you about an incident, including one at its own subprocessors. Clinics should confirm which tools are covered by a business associate agreement and on which plan; we explained that in which AI tools sign a BAA. This is general information, so take contract and regulatory questions to your own adviser.
  7. Write a one page vendor incident plan. Who reads vendor breach notices, who decides whether to disconnect, which keys get rotated, and who calls clients if needed. Keep it short enough that someone will actually follow it.

Two related items from this week's CISA list

On 18 November 2026, CISA added four actively exploited flaws to its Known Exploited Vulnerabilities catalog. Two are worth a quick check at a small regulated firm. One is in Synology DiskStation Manager, the software on many office network storage boxes, with a federal deadline of 21 November. The other is in LiteLLM, an open source gateway some developers and IT providers use to route staff requests to AI models, with a federal deadline of 2 December. The list also includes REDCap, which some research teams use for study data. Those deadlines bind US federal agencies, but they are a sensible target for anyone running these products. If you are not sure whether any of them run in your office, ask your IT provider for a yes or no in writing. (CISA, 18 November 2026)

What is still unknown

As of 20 November 2026:

  • Fireworks says the forensic investigation of the second incident is still under way. The full list of affected Fireworks customers has not been published, and more companies may yet disclose that their data was involved.
  • Notion has not published a count of affected workspaces beyond under 1%, and says its investigation is ongoing.
  • It is not public which cloud provider had the incident that led into Fireworks development machines.
  • The vendors behind the 19 November bank and card leaks have not been named.

We will update this post if Fireworks or an affected customer publishes a material change.

How we help

An AI Safety Check is the quickest way to answer these questions for your firm. We find the AI tools already in use, map which vendors and subprocessors handle your data, check sign in, app consent and retention settings, and give you a short written list of what to keep, change or switch off. In week one we also run a 60 minute staff workshop where your team lists the tasks they would hand to AI, so the review reflects how people actually work.

Because vendor risk sits on identity, devices and the network, the findings often lead to tenant and security work as well. See our cybersecurity services in San Diego and Microsoft 365 services. Rated 5.0 on Google. Book a 15 minute call to talk it through.

FAQs

What is an AI vendor data breach?

It is a breach at a company that provides AI services, such as running models, storing prompts or hosting an AI feature, that exposes data its customers sent to it. The Fireworks AI incident is an example: the breach was at Fireworks, but the data exposed belonged to customers of Notion, Uber and others. (Fireworks AI security bulletins; TechCrunch, 12 November 2026)

What is a subprocessor?

A subprocessor is an outside company that a vendor uses to handle your data on its behalf, for example the company that runs the AI model or hosts the servers. Business software vendors usually list them in a published subprocessor list or a data processing agreement.

We use Notion. Were we affected?

Notion says it has contacted the affected customers, and that if you have not heard from it, it has not identified your workspace as affected and no action is needed now. If you were contacted, treat the data as exposed and rotate any credentials that were pasted into Notion AI during the incident window, as Notion recommends. (Notion security advisory)

Should we stop using AI tools after a breach like this?

Usually not. Banning tools tends to push staff onto personal accounts the firm cannot see. A better response is to keep a short list of approved tools, on firm accounts, with known subprocessors, sensible retention settings and the ability to switch them off quickly.

What should we ask an AI vendor before using it with client or patient data?

Ask who processes the data on its behalf, where it is stored, how long prompts and outputs are kept, whether they are used for training, whether it supports single sign on with Microsoft 365, how quickly it will notify you of an incident including one at its subprocessors, and, for clinics, whether it will sign a business associate agreement on your plan.

Would our security monitoring catch a breach at a vendor?

Often it would not. Uber said it had alerts on its Fireworks environment, but they did not fire because the attackers used Fireworks' own valid credentials. That is why limiting what each vendor can reach, and being able to revoke access quickly, matters more than detection alone. (TechCrunch, updated 16 November 2026)

---

*This post is general information, not legal or compliance advice. We checked the Fireworks AI security bulletins (including the 17 November 2026 update), Notion's security advisory, TechCrunch's reports of 12 and 13 November 2026 (updated 16 November), AP and SecurityWeek of 19 November 2026, and CISA's 18 November 2026 alert on 20 November 2026. Details may change as investigations continue, so check the current position and talk to your own adviser about your firm's situation.*