Enki

Notion AI chats exposed in Fireworks AI breach

By

Published

Reporting from TechCrunch

Hackers who breached Fireworks AI got Notion AI chat transcripts from under 1% of workspaces. What to check and rotate now.

What it means for founders

Two lessons, one for Notion customers and one for anyone shipping AI features.

If your team uses Notion AI, check whether Notion has contacted your workspace owner. If it has, download the affected data and search it for credentials first. Notion's own advice is to rotate any password, API key or token found in a chat, then watch for phishing that uses details from those chats. Secrets pasted into AI tools keep turning up in incidents; Enki recently covered secrets leaking through Claude Code approvals.

If you send customer prompts to an inference vendor, this is your risk too. Notion says Fireworks had committed by contract to keep none of its data, yet its investigation found data was retained; Fireworks says its inference service keeps no data by default and was not affected. List every provider that receives user text, check what each actually stores and for how long, and ask whether a breach at their cloud provider reaches your data. Your customers will hold you responsible, not your vendor.

What we don't know yet

Neither company has said how many transcripts or users were exposed, or whether other Fireworks customers lost chat data. Fireworks has not named its cloud provider or said whether the same attackers were behind both incidents. It is not known whether the stolen data has been sold or published.

The story

Attackers who breached Fireworks AI obtained Notion AI chat transcripts, Notion has confirmed in a customer advisory. Notion says under 1% of workspaces were affected and that it has contacted those customers directly.

Fireworks runs open AI models for developers. According to TechCrunch, the company says the attacker broke into its cloud provider and from there reached a Fireworks development system holding customer data, including billing details and certain datasets. Fireworks believes the group is financially motivated and wants model access it can resell. It is Fireworks' second incident in about a month: it disclosed the first on October 9, and says attackers got in again on October 24 through a breach at a cloud provider the day before.

What Notion lost, and what it did not

Notion says it used Fireworks for two things in its product: classifying chats for its Insights feature, which shows customers how they use AI, and serving open-weight models. It also used data volunteered through its early access program to experiment with training models on Fireworks. According to TechCrunch, the stolen transcripts came from customers who used Insights.

The affected transcripts include the full text users typed into Notion AI chats, and the full text of the AI's replies. That can include names, email addresses and API tokens pasted into a chat, and fragments of pages the assistant quoted. Notion says most transcripts lacked the automatic metadata, such as user name and workspace, that would tie a chat to a person, and that AI Meeting Notes audio, transcripts and summaries were not affected unless they appeared in a chat.

Notion says it cut its systems off from Fireworks on November 3 and that the data held at Fireworks was deleted on November 4. Insights is switched off while it investigates, and Notion no longer serves open-weight models through Fireworks. The exposed data is available to workspace owners and affected members through the ? menu in the workspace sidebar, under Security incident review.

Sources

Tools in this story

We may earn a commission if you sign up through our links. It never affects our ratings or which stories we cover.

Notion AI logo
Notion AI

Your workspace, now with a brain

7.8Editor’s scoreVisit Notion AI

More in Policy & Safety

How Enki covers newsCorrectionsReport an error