Notion says Fireworks AI breach exposed AI chat transcripts from under 1% of workspaces
Transcripts behind Notion's Insights feature were taken in an attack on its inference vendor; Notion has disabled Insights and stopped serving open-weight models through Fireworks.
Update, . Notion added an update on 11 November saying it had revoked exposed Notion API keys found in affected users' messages. We also corrected the account of when the Fireworks cloud-provider incident happened, using Fireworks' own bulletin.
- What happened
- Notion disclosed that attackers who breached its AI vendor Fireworks AI accessed Notion AI chat transcripts from fewer than 1% of workspaces.
- What’s new
- Fireworks confirmed on 8 November that Notion customer data was involved, after an earlier notice that Notion found contained only synthetic test data.
- Why it matters
- Users who pasted passwords, API keys or other secrets into Notion AI chats may have exposed them through a third party, not Notion itself.
Notion has begun notifying customers that Notion AI chat transcripts were accessed in a security breach at Fireworks AI, a cloud platform it used to run open-weight models. According to Notion's incident page, fewer than 1% of workspaces are affected. Notion stresses that its own systems were not breached.
What was taken and how traceable it is
The exposed data came from the dataset behind Insights, a feature that shows workspace administrators how their teams use AI. Notion used Fireworks for classification jobs for Insights and as a general inference provider for open-weight models. Notion says that for the vast majority of affected workspaces the transcripts were unattributable, meaning the chats weren't labeled with an author or workspace and contained no tool call results. That wording leaves open that some transcripts could be tied to their source, and SecurityWeek notes Notion has not said how many customers were affected. BleepingComputer reports that Notion's FAQ says Fireworks had contractually committed to zero data retention for Notion data.
Notion has disabled Insights during the investigation and no longer serves open-weight models through Fireworks. Affected customers have been emailed instructions to retrieve and review the data inside their workspace; BleepingComputer reports this happens through a Security incident review option in the web interface. "We recommend treating this data as exposed," Notion says, advising anyone who pasted passwords, API keys or other credentials into Notion AI during the incident to rotate them. Notion says users who have not been contacted need take no action for now. On 11 November Notion added an update to its page saying it had revoked exposed Notion API keys found in affected users' messages during its investigation.
A timeline the reports tell differently
Notion's own account has two notices. On the evening of 2 November, Fireworks told Notion a small dataset had been accessed; Notion found it held synthetic Notion AI test data and no customer data, but by the morning of 3 November it had disconnected from Fireworks and deleted all its data there. On 6 November, Fireworks warned that customer data may have been accessed, confirmed the affected dataset on 8 November, and Notion contacted affected customers on 9 November. SecurityWeek's account differs, saying Fireworks told Notion on 2 November that Notion AI data had been accessed. BleepingComputer's version omits the 6 November warning and moves from the 2 November alert to the 8 November confirmation.
Notion revised its page on 10 November, saying the changes were for style and clarity and that the facts had not changed.
Two breaches at Fireworks
Both outlets report that Fireworks was breached twice in October. The earlier intrusion involved credentials: SecurityWeek describes compromised internal credentials, while BleepingComputer says attackers reached environment variables holding API keys and tokens for a limited number of accounts. The incident that affected Notion came through a third-party cloud provider hosting Fireworks development machines. Fireworks' own bulletin, updated on 9 November, says the provider had a security incident on 23 October and that the attacker used it to reach Fireworks development machines on 24 October. BleepingComputer, which dates the access to 23 October, says resource metadata, billing information, model weights and datasets of some customers were compromised; SecurityWeek dates the attack to 24 October, says attackers ran an unauthorized workload, and links it to a rise in LLMJacking activity.