Skip to content
NewsIncident

A Fireworks AI breach spilled Notion's AI chat transcripts into attackers' hands

· by

The short version

Notion disclosed that a security incident at AI infrastructure vendor Fireworks exposed chat transcripts from its Insights feature, with Fireworks confirming the final affected dataset on November 8.

A Fireworks AI breach spilled Notion's AI chat transcripts into attackers' hands

Notion disabled its AI-powered Insights feature and cut off its inference vendor Fireworks AI after confirming that a breach at Fireworks exposed chat transcripts tied to that feature, per Notion's security advisory.

What

The timeline Notion published runs over a week. Fireworks first flagged unusual access on November 2, but that dataset turned out to be synthetic test data, not customer content. Notion still disconnected its systems from Fireworks by November 3 and deleted its data on the Fireworks side by November 4, per the advisory. On November 6, Fireworks came back and said malicious actors may have reached actual customer data. Fireworks confirmed the final scope of that dataset on November 8, and Notion's own review found it included Insights chat transcripts. Notion says fewer than 1% of its workspaces were affected and started notifying those customers on November 9. The root cause traces to a breach at the cloud provider hosting Fireworks' research-and-development machines, which Fireworks disclosed separately in late October, per BleepingComputer.

Two Fireworks jobs sat behind Notion's product: classification work behind Insights, which summarizes a workspace's own AI usage, and general inference for open-weight models elsewhere in the app. Both are off now. Most of the exposed transcripts are "unattributable," per Notion, meaning the text carries no label tying it back to a specific user, workspace, or company, though the chat content itself is still exposed if it contained sensitive material.

Insights stays off until the investigation closes

Notion has not said when Insights will return, only that it is disabled "as we investigate." Workspace admins get one concrete action from the advisory: if anyone pasted passwords, API keys, or other credentials into Notion AI during the incident window, rotate them now, because exposed content counts as compromised regardless of whether it traces back to a person. Affected workspace owners also get a CSV export of exactly what was exposed, pulled from the workspace sidebar, which is a more specific remediation step than most vendor breach notices offer.

A SaaS company running a user-facing AI feature on a third-party inference vendor has little visibility into that vendor's own infrastructure. When Fireworks' systems were compromised, the breach surfaced inside Notion's product with no advance warning to Notion's own users. Any tool that routes AI features through a shared inference layer carries that same exposure; this incident shows what it looks like when it lands.

What to watch next

The advisory page is still being updated as the investigation continues, so the workspace count and the "unattributable" framing could move. Fireworks has not published a detailed post-mortem naming the cloud provider involved, and whether other products that piped AI features through Fireworks disclose their own exposure is the next signal worth tracking.

Sources