AI & Automation

An AI Model Provider Was Breached and Customer Chat Transcripts Leaked: Vetting AI Vendors

November 13, 2026 · 5 min read · Etzal Group Team

On 12 November 2026 TechCrunch reported that Fireworks AI, a platform providing access to open AI models, confirmed a cyberattack that exposed data belonging to some of its customers. One of those customers, Notion, said chat transcripts from users of its AI feature were stolen, affecting about 1% of Notion's customers.

For a mid-sized company the lesson is not about those two names. The applications you use or build very likely pass data to a model provider you never chose yourself, and that provider has suppliers of its own.

What was reported

According to the TechCrunch report:

  • The attackers breached the cloud provider Fireworks uses in order to reach Fireworks' development systems. That cloud provider is not named.
  • According to Fireworks' latest security advisory, the attackers took customer billing information and some datasets.
  • Fireworks says it has tens of thousands of customers worldwide, including Harvey and Cursor. The source names those two only as customers, not as affected parties.
  • Notion said the breach affected about 1% of its customers. On 3 November Notion disconnected its systems from Fireworks and deleted the data stored there.
  • What was taken from Notion: chat transcripts of Notion AI customers who used the "Insights" feature, covering "any input to or output from the model that was written to the chat". Notion said most of the data could not be linked to specific customers.
  • Fireworks believes the hackers were financially motivated and wanted to sell LLM tokens obtained through unauthorised access.
  • This is Fireworks' second incident in two months, following one in October 2026.

The incident is still developing and the figures may change. Check the latest official statements from the companies involved before quoting it in internal documents.

The chain you cannot see

Lay the event out as a chain: a user types into an app, the app forwards the text to a model provider, the model provider runs on a cloud provider. The break-in happened at the far end of that chain, and what leaked was chat content from the near end.

End users hold a contract only with the app. They did not choose the model provider, let alone the cloud provider behind it. That is the heart of an AI vendor security breach: the risk sits two or three layers beneath the contract you actually signed.

Note also what was stolen. Chat transcripts contain whatever users typed: contract excerpts, sales figures, customer complaints. Nobody classifies that text before it is sent.

Where you sit in the chain

There are two positions, and many companies occupy both.

You use applications with AI features. You need to know where the vendor forwards your data.

You build AI features for customers or staff. You are the one choosing the model provider, and your customers will ask you the same questions.

An AI supplier map on one sheet

Make a table with one row per application that uses AI. The columns:

  1. The application and the AI features switched on.
  2. The model provider behind it. Look in the vendor's subprocessor list, or ask.
  3. What data is sent: only the user's question, or documents and context too.
  4. What the model provider stores, and for how long.
  5. Whether the data is used to train models.
  6. Who tells you about an incident, and how fast.
  7. How to switch the feature off, and who has the authority.

Column four matters most. Data that is not stored cannot be stolen from storage. Ask whether request logging can be disabled or shortened, and whether that requires a particular plan.

Column seven is often blank. Notion cut the connection and deleted data on 3 November. Ask yourself: if your model provider were breached tonight, how long would it take you to disconnect, and would the application still run without its AI feature?

Clauses worth asking for

Indonesian law offers a handhold here. According to RSM Indonesia's summary of Government Regulation (Peraturan Pemerintah, or PP) 33/2026, which takes effect on 16 January 2027, a processor needs the controller's written approval before engaging a subprocessor, and must report a data protection failure to the controller at the first opportunity. We take this from a consultancy summary, not from the official text of the regulation.

Translated into a contract with an application vendor:

  • A named subprocessor list, including model providers, with a duty to notify you before it changes.
  • An incident notification deadline stated in hours, covering incidents at subprocessors.
  • A retention limit for model inputs and outputs, and no use for training without consent.
  • A right to request deletion, with evidence that it was done.
  • A right to switch off AI features without losing the application's core functions.

The benefit is clear. There is also a cost: large vendors rarely amend standard contracts for mid-sized customers. Where a clause cannot be negotiated, use the settings that exist, such as disabling the feature for certain user groups, and record that decision.

Legal obligations differ from company to company. Confirm your own case with legal counsel or the data protection authority.

If you are the one building

  • Send as little as possible. Remove or mask personal data before text reaches the model provider when the task does not need it.
  • Choose the shortest retention setting the provider offers, and keep the evidence.
  • Keep API keys somewhere they can be rotated in minutes. The motive Fireworks described, selling tokens obtained through unauthorised access, shows that model access itself has resale value.
  • Design so the model provider can be replaced. Depending on a single provider that has had two incidents in two months is an operational risk, not only a security one.
  • Draft the customer message for an incident in advance. It is far easier to write when things are calm.

This week

  1. List the applications with AI features switched on across the company. Start with note-taking tools, CRM and customer service.
  2. For the three holding the most sensitive data, fill in the seven columns above.
  3. Send each vendor one email: who is your model provider, and what is the retention policy?
  4. Tell staff what must never be pasted into any AI chat.

If you are building an AI feature and want its supply chain reviewed, Etzal Group's AI consulting team can help.

Sources

Frequently asked questions

What happened at Fireworks AI?

According to TechCrunch on 12 November 2026, Fireworks AI confirmed a cyberattack. Attackers breached its cloud provider to reach development systems, then took customer billing information and some datasets.

Why were Notion customers affected?

Notion's AI feature used Fireworks. Notion said chat transcripts from users of its Insights feature were stolen, affecting about 1% of its customers.

What should I ask an app vendor that offers AI features?

Which model provider sits behind the feature, what data is stored there and for how long, and how quickly you will be told if an incident occurs.

Author

Etzal Group Team

Etzal Group Team, helping businesses grow with technology. Contact us.

Related service

WhatsApp AI Chatbot

We build, train, and run a WhatsApp AI agent that answers customers 24/7, qualifies leads, and follows up.