Fireworks AI attack exposed Notion customer chat transcripts
A cyberattack on AI infrastructure provider Fireworks AI exposed a dataset containing Notion AI chat transcripts, according to reporting by TechCrunch and follow-up disclosures from Notion. Notion says fewer than 1% of its workspaces may have been affected. The incident did not result from a breach of Notion's own systems, but from data that Notion had sent to Fireworks for processing.
The transcripts were used for Notion's Insights feature, which helps workspace administrators understand how teams use AI. Notion says the affected dataset was confirmed by Fireworks on 8 November after an initial alert on 2 November had been assessed as involving synthetic test data. Notion disconnected its systems from Fireworks, deleted its data there and disabled Insights while it investigated. It also stopped serving open-weight models through Fireworks.
For most affected workspaces, Notion describes the transcripts as unattributable: they were not labelled with the author or workspace and did not include tool-call results. That limits the available context, but it does not make the text harmless. Prompts and answers can still contain private business information, personal details, passwords or API keys. Notion has contacted affected customers and recommends that they review the data through the workspace security process and rotate credentials that may have been pasted into an affected chat.
Fireworks' own 9 November bulletin describes a separate chain of activity involving a third-party cloud provider that hosted its research and development machines. Fireworks says an unauthorised workload reached those machines on 24 October and that its investigation found impacts involving resource metadata, billing information and, for a limited set of customers, secrets, datasets and incomplete model weights. It says its inference and training services are zero-data-retention by default, while stored datasets and other internal systems required additional investigation.
The practical lesson for AI users and companies is about the difference between a vendor's retention promise and the places where data is actually stored for batch jobs, analytics or testing. Before sending sensitive prompts to an AI feature, teams should map the providers behind it, ask which functions retain content, restrict secrets in prompts and prepare a rotation plan. The incident also shows why third-party risk remains central even when the primary application itself was not compromised.