A breach at model-platform provider Fireworks AI allowed attackers to access customer billing information, datasets and some Notion AI chat transcripts. Fireworks said the intruder first compromised its cloud provider and then reached a development system containing customer data.
Notion says the incident affected about 1% of its customers who used the Insights feature. Stolen transcripts could include any prompt or model output written into those chats. The company said much of the material could not be attributed to a particular customer, but that does not remove the sensitivity of the conversation content.
After Fireworks alerted it, Notion disconnected its systems on November 3, deleted data held by the provider and stopped all connectivity. Fireworks believes the attackers were financially motivated and primarily sought model-access tokens that could be resold for unauthorized usage.
This is the second disclosed Fireworks security incident in two months. The event also shows how an application can inherit risk from the model host and that host’s own cloud suppliers. Organizations using external inference services need retention limits, scoped credentials, incident cutoffs and an inventory of which prompts and outputs each provider stores; otherwise a breach several layers down can expose end-user conversations.