On November 9, Notion’s Chief Information Security Officer emailed me. My workspace was among fewer than 1% of Notion workspaces identified as affected by a security incident involving Fireworks AI, an infrastructure provider behind Notion AI. That percentage refers to workspaces, not users. The total affected count has not been publicly established. [1]
I reviewed the affected-content export Notion provided. It contained 123 message rows across eight chat threads. I didn’t identify recognizable password or API-key patterns in my bounded review, but that isn’t a forensic guarantee. What I could confirm was that some of my AI conversations had been exposed.
That raises a question beyond my own workspace: when a business uses an AI product, does it know who else handles its data and how those providers’ controls are verified?
The supplier chain behind one feature
Notion used Fireworks for classification jobs supporting Insights, a feature that helped administrators understand Notion AI usage. Fireworks also provided inference infrastructure for some open-weight models.
Notion says it disconnected Fireworks on November 3, deleted its data from Fireworks on November 4, and contacted affected customers on November 9 as the investigation developed. [1]
The exposed data included chat inputs and assistant responses. Most affected transcripts lacked metadata explicitly identifying a user or workspace. But conversation text can reveal names, projects, and business context without an account label. [2]
This was a confidentiality incident in a downstream provider, not established evidence of an intrusion into Notion’s underlying workspace databases.
Two incidents, unresolved attribution
On November 17, Fireworks distinguished two separate security incidents.
The first, spanning August 3 through October 7, involved compromised credentials and a vulnerability in a Fireworks service. The second, spanning October 23 through October 30, involved a compromised third-party cloud provider and access to Fireworks development and production systems.
Fireworks reported both incidents contained and remediated. Its forensic investigation of the first had concluded, while a separate investigation of the second was underway as of that disclosure. [3]
The public record does not establish which incident exposed the Notion Insights transcripts. Attributing that exposure to either attack path would go beyond the available evidence.
What did zero retention actually mean?
Notion’s incident FAQ makes an important distinction. Its enterprise zero-data-retention commitment for generating AI responses did not apply to features intentionally storing conversations, including Insights.
But Notion also says Fireworks separately committed to zero retention for the Insights analysis sent to its platform. During the investigation, Notion discovered that data had nevertheless been retained at the Fireworks endpoints it used. [2]
The contractual assurance described what should happen. The observed data handling demonstrated something different.
The public record still does not explain why the data persisted or what endpoint-level evidence would independently establish that the relevant information had been deleted.
Trust should be verifiable
Third-party infrastructure is not inherently a problem. Much of modern software depends on it.
But organizations should be able to identify which providers process their data, which features retain it, how retention and deletion are tested, and what records are available when something fails.
NIST’s voluntary AI Risk Management Framework addresses these third-party supply-chain risks under GOVERN 6, including contingency planning for supplier incidents. [4]
An AI product can be useful, save time, and create commercial value without providing adequate visibility into its underlying controls.
A useful product deserves a verifiable supply chain.
A contract tells customers what should happen. Operating evidence shows what actually did.