Anything typed into a chat box can end up somewhere unintended, and for a slice of Notion’s customers that turned out to include passwords and API keys pasted into the workspace assistant. Notion says the exposure came not from its own servers but from an outside AI infrastructure company, Fireworks AI, which was breached in the weeks before Nov. 9.
Read next on Morning Overview: Two Kentucky residents died when a small plane crashed at Missouri’s Lake of the Ozarks, and the NTSB is investigating
Notion’s advisory puts the reach at fewer than 1% of workspaces. For a product with millions of users, that is still a large number of conversations, and the contents are the part that worries security teams: full chat messages and assistant replies that the company says may include names, email addresses and credentials.
A vendor Notion relied on for Insights and model hosting
According to Notion’s help-center notice, Fireworks AI served as a third-party AI infrastructure vendor. Notion used it for Insights classification jobs, as an inference provider for open-weight models and, through its Learning and Early Access Program, to experiment with training task-specific models.
The contract carried a zero-data-retention commitment for Notion’s platform use. The notice says Notion later learned that Fireworks had in fact kept data from the endpoints Notion used. That mismatch between the promise and the practice is the center of the story, because customers had no direct relationship with the vendor and no way to check the retention claim themselves.
The advisory first published Nov. 9 gives the sequence. On the evening of Nov. 2, Fireworks told Notion a small dataset of synthetic Notion AI testing data had been accessed. Notion disconnected from Fireworks the next morning, deleted data in the vendor’s system and stopped communicating with it. On Nov. 6, Fireworks said malicious actors might also have reached potential customer data, and on Nov. 8 it confirmed the new dataset, which included Insights chat transcripts.
Contents of the exposed AI chat transcripts
The affected material is the full text of Notion AI chat messages and the assistant’s outputs, which may carry fragments of pages or databases the assistant referenced, plus possibly text selected on the current page. Most affected transcripts lacked metadata linking them to a specific user, workspace or company, so an attacker holding them sees conversation content without a clear label of whose it is.
Notion lists what was not touched: most tool-use data, chat attachments, the rest of page content, and AI Meeting Notes unless a user typed them into a chat or the assistant quoted them. The company’s guidance, relayed in the AI Brief of Nov. 13, is that the exposed chats from Insights classification may contain credentials.
The unlabeled nature of the data cuts both ways. It lowers the chance that a stranger can link a stolen API key to its owner, yet a key sitting in plain text in a chat can still be tried against the services it belongs to.
Notion’s response and the guidance for workspace owners
TechCrunch reported on Nov. 12 that the cyberattack on Fireworks AI let hackers reach Notion customers’ chat transcripts. Notion has disabled the Insights feature during its investigation, stopped serving open-weight models through Fireworks and begun auditing critical vendors to confirm zero data retention and data minimization.
The company contacted affected customers directly. Those who were contacted can download the affected data from the workspace sidebar under the question-mark menu’s security incident review. Notion tells them to rotate any exposed passwords, API keys or tokens, starting with the riskiest, and to watch for phishing that mentions names, emails or the incident. Workspaces that were not contacted are told no action is needed now, while the investigation continues.
Notion’s current advisory page states that the investigation is ongoing and that the company will update the page as it learns more. A workspace owner who has not been contacted is not guaranteed to stay in that category, since the notice says Notion will reach out to any workspace it later identifies as affected, so the list of affected customers could still grow.
Fireworks’ own two incidents
The Notion exposure sits inside a longer chain at the vendor. Fireworks first detected unauthorized access on Oct. 6 and disclosed it Oct. 9; attackers had reached environment variables for a limited number of customer accounts, according to eSecurity Planet’s account. A second intrusion followed after a cloud provider hosting Fireworks research machines reported its own incident on Oct. 23, and Fireworks published a bulletin on Nov. 9.
Fireworks said it engaged Palo Alto Networks’ Unit 42 and notified law enforcement, and attributes the activity to a financially motivated group interested in so-called LLM-jacking. It says it does not store payment card numbers. The same report notes that Render’s chief executive, Anurag Goel, posted a Notion notification on X confirming Render was among the affected customers, evidence that the exposure reached business users.
Whether Fireworks’ other disclosed material, including billing addresses and partial model weights, overlaps with the Notion transcripts is not stated in the sources; the question is still open as the investigations continue.
This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.
More from Morning Overview
- Shaq spent $1.3 million on three Rolls-Royce Phantoms to prove a point to a salesman
- The Justice Department asked courts to undo approvals for three offshore wind farms, one already making power
- Over 20 trillion gallons of rain fell on the Four Corners states and Nevada in two weeks
- 11 engines built to run well past 200,000 miles