Notion AI Fireworks Breach: Review Your Chats in 2026
Read the affected conversations without mistaking later review annotations for leaked data.
November 10, 2026

Notion's current incident notice says the Notion AI Fireworks breach involved chat transcripts. Fewer than 1% of workspaces may be affected, and no other workspace data was affected. Workspace owners should use Notion's official incident review tools and check the conversations themselves. Missing account labels say little about what a conversation reveals.[1]
That distinction matters because an unlabeled conversation might still contain a client's name, an unfinished personnel decision, or a sentence someone dictated instead of typing. These are examples of possible content, not findings about the exposed chats.
November 10 attention, October incident dates
This disclosure is recent, but it doesn't show that an attack began today. In its November 9 update, Fireworks said an incident at a cloud provider occurred on October 23, followed by unauthorized access to Fireworks development machines on October 24. The investigation is still ongoing.[2]
BleepingComputer separately reported the breach and listed affected categories, including some datasets and incomplete model weights. We read the article in a Wayback snapshot captured November 10. Its report supports the broader Fireworks incident, but does not verify the contents of any individual Notion workspace.[3]
On November 10, we found the Notion discussion on Hacker News' front page. The comments are reactions, not a replacement for the notice or your workspace's review material.[4]
Notion says the transcripts that may have been exposed had no user or workspace labels and contained no tool-call results. It disabled Insights and disconnected Fireworks. The company advises treating the affected data as exposed and rotating any credentials pasted into those conversations.[1]
The notice gives affected owners a clear way to inspect the material, which helps. Still, colleagues should be careful about using "unattributable" to sum up the story. Removing a label doesn't change what the words underneath mean.
Read what the conversation reveals on its own
A fictional events company asks an AI chat to revise a speaker invitation. The request includes a replacement venue, a cancellation that hasn't been announced, and a guest's accessibility requirements. No one shares a password, and the invitation never becomes public.
The reviewer still needs to ask whether the guest could recognize the description and whether the cancellation names a business partner. They also need to check which details appeared in the prompt and which appeared only in a separate document. Keep those distinctions in the review instead of assuming every related document was exposed.
A summary can give the wrong impression. "No account identifiers" may be accurate for one field, but it doesn't capture what the conversation reveals. In that fictional invitation, a name in the message would still matter, even if the export had no author column.
When you can, read the text itself. A label like "marketing" says little about how sensitive a paragraph is. Even a short scheduling request might mention a detail deliberately left out of the final invitation.
Don't fill in missing details. If the reviewed transcript only shows a request for an invitation, don't list a related contract as exposed just because you remember working on both that afternoon. Note what's unclear and ask the person managing the investigation.
Separate the exposed text from review annotations
Notion's Enterprise review export can include added user metadata, but the notice says those annotations weren't exposed. Access and export options depend on your plan and workspace role. Follow the current owner instructions on desktop.[1]
Spreadsheets can make it easy to confuse the two. Before copying rows into a working document, label the fields that came from the affected material and those added during the investigation. Keep the original export unchanged in an approved restricted location.
Suppose a review row links an unlabeled chat to a fictional employee named Dana. Someone may need to include Dana's name in an internal note so they can ask about the conversation. The note should make clear that Dana's name came from a later annotation. Otherwise, the next person who reads it may think the attacker also linked the chat to Dana.
An AI summary could combine the two fields into one confident-sounding sentence. Before sharing it, compare the wording with the export. If it says, "the leaked file identified Dana," check which original field supports that claim. Correct it if the link appears only in the review copy.
Keep the original CSV unchanged and work from a copy. Save enough row details for an authorized reviewer to find each source. An untraceable summary will waste time when the next update arrives.
Make a short review record without spreading the chats
For each conversation you review, note the business task and any sensitive details. Include the source row. Keep proposed actions separate from completed actions, especially when someone else needs to revoke a credential or contact an affected colleague.
Share only what the next person needs. For an authorized teammate, "Check the venue-change paragraph in row 14" may be enough. Posting the whole conversation in a general Slack channel gives people who don't need it another copy.
If the text contains a credential, follow Notion's rotation recommendation and your organization's process.[1] Record who handles it and when they confirm it's done. Rewriting the chat or deleting a local export doesn't prove the credential was revoked.
You can say when you don't know something. You may be able to see your own conversation, but not the owner's broader review. Be clear about that. Don't treat limited access as evidence about the whole company, or use a public commenter's experience to fill in what you can't see.
Before sending an update, check the official notice again. Add the review date beside your conclusions so colleagues know which version you relied on. If a later notice changes the scope, revise your conclusion and keep the earlier record under your usual incident procedures.
Dictation can help write the handoff, not resolve exposure
On a desktop, hold a shortcut in DictaFlow and speak. Release it to type the words at your cursor.[5] You can write a short review note in an approved destination. DictaFlow doesn't inspect Notion's incident export or undo a conversation already sent to another service.
Our Notion dictation guide explains text entry separately. Changing how you enter text still doesn't show which past chats were affected.
For now, share the workspace's review materials with the person handling the response. Include the relevant rows, and mark any notes added later so they're separate from the original content. If you can't access the material, ask the workspace owner for the official review route. Don't paste a sensitive conversation into another chat to get an opinion.
Sources
[2] Fireworks security bulletin, November 9
[3] BleepingComputer, archived November 10, 2026 at 15:20:34 UTC