Notion says some transcripts from Notion AI chats may have been exposed through its inference provider, Fireworks AI. Following its investigation, Notion estimates that fewer than 1% of its workspaces may be affected. It has contacted affected workspace owners and offers them a way to review the data.

Evidence status: vendor-confirmed data in an incident dataset, with individual impact still under review. Notion says Fireworks notified it on November 6 of potential access to customer data, then confirmed a new dataset on November 8. Notion found chat transcripts in that dataset. The notice does not establish that every possibly affected conversation was read by an attacker.

What the data included

Notion says the transcripts came from Insights, a feature that classified customers’ AI usage. The affected chats carried no user or workspace labels and no results from tool calls, such as retrieved page content. The words people typed into a chat can still identify a person or organization, or contain a secret. “Unattributable” describes missing labels, not guaranteed anonymity of the conversation text.

Notion says no other workspace data was affected. It disconnected from Fireworks on November 3, disabled Insights and stopped serving open-weight models through that provider. Fireworks’s broader incident update describes unauthorized access through a third-party cloud provider in October and says its investigation is ongoing. Its account of affected data categories and Notion’s account of chat transcripts should be read separately; neither establishes the entire scope of the other’s systems.

Review transcripts and rotate exposed credentials

Notion directs affected owners to their workspace sidebar, then ? → Security incident review, using a current desktop browser. Owners of workspaces with a verified domain can download the affected-content CSV and an annotated version to assess who sent which chat. Notion says those annotations were not part of the exposed dataset. All affected owners can download an affected-members CSV and share a review link so members can see their own data.

Update: November 10, 21:44 EST — clarified who can download affected chats

Notion’s current help page says the affected-content CSVs are available to owners with a verified workspace domain. This corrects our earlier description of eligibility as “Enterprise owners with managed members.” All affected owners can download the affected-members list and share a per-member review link. Notion also says the transcript includes both input and output written to a chat; AI Meeting Notes data is outside this incident unless included in a Notion AI chat. This is a correction to review guidance, not evidence of more affected workspaces.

If a chat contains a password, API key or other credential, Notion recommends rotating it. Review the affected transcripts before deciding which other sensitive information needs response. Notion says owners who have not heard from the company have no action to take right now, based on its current investigation.

What the thread adds

In the Hacker News discussion, Shank questions whether transcripts without workspace labels can be reliably attributed for notification. _fw points out that the content itself can identify people even without metadata. Those are readers’ interpretations, not evidence that additional workspaces were affected. Notion’s notice offers affected owners their data for review and says it contacted them directly.

From a 49-comment thread, in HN’s own order.

Caveat: Notion and Fireworks are describing an ongoing investigation; no independent forensic report of the Notion transcript dataset is available in these notices.